CHOOSE YOUR AGENT INFRASTRUCTURE
Run the sessions.
Own the infrastructure.
Put Claude, Codex, Antigravity, Pi and 23 more agent CLIs in persistent, isolated workspaces on your servers. devctrl prepares the containers, connects project resources and gives each developer a private environment — with sessions and operations in one console.
AT A GLANCE
Key features
Choose a product to compare hosting, isolation, files, teamwork and operations.
- Yes
- Limited
- With extension
- Manual setup
- Unavailable / unverified
78 unavailable or unverified features
Amika comparison
Agent VM workstations on your machines or cloud accounts with private connectivity. Current CLI operations use the hosted control plane; a complete supported self-hosted control plane is not established.
Sources: Amika: own-compute VMs and private networking · Amika: open components, hosted control plane and Rig interfaces · Amika: removed local CLI mode and current self-hosting limitations| Feature | devctrlSelf-hosted control plane | AmikaVM workstations + hosted control plane |
|---|---|---|
| Software pricing | Free self-hosted software | Website advertises free own-compute use for teams of up to three during early access; cloud costs and broader plans require current terms |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | Not documentedNot documented in reviewed sources |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesVM workstations on your own machines and cloud accounts |
| Run the control plane on your server | YesRun controller and browser gateway on your server | LimitedCurrent Rig CLI talks to hosted control plane; local Docker CLI mode was removed. Self-hosted gateway path remains described as upcoming |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedNot documented in reviewed sources |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedNot documented in reviewed sources |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | VM-based Rig workstations; several agents can share a Rig |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedNot documented in reviewed sources |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Not documented in reviewed sources |
All features
| Feature | devctrlSelf-hosted control plane | AmikaVM workstations + hosted control plane |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Website advertises free own-compute use for teams of up to three during early access; cloud costs and broader plans require current terms |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Not documented in reviewed sources |
| Execution compute costs | Your local/remote compute bill | Not documented in reviewed sources |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | Not documented in reviewed sources |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | Apache-2.0 CLI/daemon/SDK components; README lists control plane, UI and relay as closed |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Not documented in reviewed sources |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Your compute or hosted devboxes; current control-plane service remains hosted |
| Run agents on this computer | YesStart sessions on this computer; Docker required | LimitedOwn-machine VM provider is documented; removed --local Docker CLI path is not supported |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | YesOptional Amika-hosted devboxes |
| Run the control plane on your server | YesRun controller and browser gateway on your server | LimitedCurrent Rig CLI talks to hosted control plane; local Docker CLI mode was removed. Self-hosted gateway path remains described as upcoming |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | VM-based Rig workstations; several agents can share a Rig |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | Not documentedNot documented in reviewed sources |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | Not documentedNot documented in reviewed sources |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | Not documentedNot documented in reviewed sources |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesVM workstations on your own machines and cloud accounts |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | Not documentedNot documented in reviewed sources |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedNot documented in reviewed sources |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedNot documented in reviewed sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedNot documented in reviewed sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedNot documented in reviewed sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedNot documented in reviewed sources |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedNot documented in reviewed sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedNot documented in reviewed sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedNot documented in reviewed sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedNot documented in reviewed sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedA separate frozen-session view with retained terminal snapshots and this S3 recovery workflow is not established by the reviewed sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedAn equivalent tree of agent-reported tasks and hook-recorded native steps for each session is not established by reviewed sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedAn owner-scoped inbox in which separate agent sessions exchange messages through a managed MCP tool is not established by reviewed sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedNot documented in reviewed sources |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documentedNot documented in reviewed sources |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Not documentedNot documented in reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedNot documented in reviewed sources |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | YesRigs may be short-lived or persistent |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | Not documentedNot documented in reviewed sources |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | Not documentedNot documented in reviewed sources |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | LimitedPrivate network and HTTPS previews are documented; exact devctrl-style project/service grant equivalence is unverified |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | Not documentedNot documented in reviewed sources |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedNot documented in reviewed sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedNot documented in reviewed sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | Not documentedNot documented in reviewed sources |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | Not documentedNot documented in reviewed sources |
| Continue on another subscription | YesEligible account + saved context | Not documentedNot documented in reviewed sources |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | Not documentedNot documented in reviewed sources |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | Not documentedNot documented in reviewed sources |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedNot documented in reviewed sources |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedNot documented in reviewed sources |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedNot documented in reviewed sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedThis comparison has not established equivalent per-account native model catalogs and default selection from the cited sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedShared per-user account display ordering across clients is not established by the cited sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | YesRig/agent API, CLI and SDK interfaces |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | Not documentedNot documented in reviewed sources |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedNot documented in reviewed sources |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | Not documentedNot documented in reviewed sources |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | Not documentedNot documented in reviewed sources |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | Not documentedNot documented in reviewed sources |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedNot documented in reviewed sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | Not documentedNot documented in reviewed sources |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | Not documentedNot documented in reviewed sources |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedNot documented in reviewed sources |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | Not documentedNot documented in reviewed sources |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | Not documentedNot documented in reviewed sources |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documentedNot documented in reviewed sources |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | Not documentedNot documented in reviewed sources |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | Not documentedNot documented in reviewed sources |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | Not documentedNot documented in reviewed sources |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Not documentedNot documented in reviewed sources |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Not documentedNot documented in reviewed sources |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Not documentedNot documented in reviewed sources |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | Not documentedNot documented in reviewed sources |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedAn integrated per-container journal covering these tool classes and verified installation outcomes is not established by the cited sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedDeletion triggered by all coding sessions exiting is not established by reviewed sources; timeouts and inactivity policies are different |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | Not documentedNot documented in reviewed sources |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | Not documented in reviewed sources |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | Not documentedNot documented in reviewed sources |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedNot documented in reviewed sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedNot documented in reviewed sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | Not documentedNot documented in reviewed sources |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedNot documented in reviewed sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | Not documented in reviewed sources |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Not documented in reviewed sources |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | Not documentedNot documented in reviewed sources |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | Not documentedNot documented in reviewed sources |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedNot documented in reviewed sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | Not documentedNot documented in reviewed sources |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | Not documented in reviewed sources |
Anvaya comparison
A pre-1.0 local coding agent harness with a persistent project memory. It runs its own agents against local or bring-your-own-key models on one Linux or macOS machine; it is not a controller for Claude or Codex sessions.
Sources: Anvaya: Agentic coding harness with an experience store · Anvaya: Concurrency · Anvaya: Private source access · Anvaya: Providers and models| Feature | devctrlSelf-hosted control plane | AnvayaLocal coding harness + memory daemon |
|---|---|---|
| Software pricing | Free self-hosted software | Pricing is not set; pre-1.0 public beta with source access through a waitlist |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | LimitedSwarm runs use a detached supervisor; interactive sessions surviving a client disconnect are not established |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run the control plane on your server | YesRun controller and browser gateway on your server | Not documentedA shared server control plane is not established; the memory daemon serves local clients |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedTerminal UI and headless CLI; a browser client is not established |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Host processes under a declarative tool policy; opt-in Seatbelt/Landlock; swarm agents in shared, worktree or cloned workspaces |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Approval and confinement are separate; --yolo never widens the policy; secrets are stripped from child environments |
All features
| Feature | devctrlSelf-hosted control plane | AnvayaLocal coding harness + memory daemon |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Pricing is not set; pre-1.0 public beta with source access through a waitlist |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Local Ollama models need no key; Anthropic and OpenAI-compatible endpoints use your own API keys |
| Execution compute costs | Your local/remote compute bill | Runs on your own machine; local models need enough RAM |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | Linux / macOS terminal UI and headless CLI; Windows is not supported |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | Rust CLI harness plus a local Mind memory daemon over Unix sockets; built from early-access source |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Own coding agent harness with a persistent project memory graph; not a controller for other agent CLIs |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Local CLI and daemon on one machine; a server deployment is not documented |
| Run agents on this computer | YesStart sessions on this computer; Docker required | YesRun the anv harness on this computer |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | Not documentedRemote providers supply model inference only; vendor-managed execution compute is not established |
| Run the control plane on your server | YesRun controller and browser gateway on your server | Not documentedA shared server control plane is not established; the memory daemon serves local clients |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Host processes under a declarative tool policy; opt-in Seatbelt/Landlock; swarm agents in shared, worktree or cloned workspaces |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | YesSwarm placement gives agents separate Git worktrees or clones |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | LimitedSwarm runs use a detached supervisor; interactive sessions surviving a client disconnect are not established |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | LimitedInline approval cards and stalled-subagent heartbeats in the terminal UI |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedTerminal UI and headless CLI; a browser client is not established |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | YesLive subagent panes with heartbeats; up to four run concurrently |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedA context summarizer runs at 80% context use; compaction after completed work is not established |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | LimitedProvider keys are stripped from child environments unless a variable is explicitly allowed |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | Not documentedThe swarm board shows per-agent tokens; provider subscription quota and resets are not established |
| Continue on another subscription | YesEligible account + saved context | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | Not documentedModel providers can change without losing the memory graph; Claude, Codex or Antigravity session transfer is not established |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | LimitedChoose models per project or task manually; no delegated-task model advisor is established |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | LimitedMind exposes 14 MCP memory tools to other clients; session or resource management tools are not established |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | LimitedSubagents and swarms run Anvaya’s own harness; other agent CLIs share its memory only over MCP |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | YesBuild one static Rust binary from early-access source |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | Each swarm agent is a supervised process with heartbeats; Mind is a separate local daemon |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | LimitedSwarm status shows per-agent memory use and tokens |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | The swarm board shows per-agent tokens; cost reporting is not established |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | LimitedSwarm tasks must pass their verification command before reaching an integration branch; work is never pushed to a remote |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Approval and confinement are separate; --yolo never widens the policy; secrets are stripped from child environments |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | LimitedmacOS Seatbelt and Linux Landlock can be enabled; container or VM isolation is not established |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | LimitedThe strict profile denies network; other profiles gate fetch and search tools through the same policy |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | YesDeclarative read/write policy checked for every tool call; kernel enforcement is opt-in |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | A tool-layer policy enforces reads, writes, execution and network beyond instructions |
boxes.dev comparison
A hosted service that gives each Claude or Codex agent its own cloud Linux machine started from a snapshot of your dev environment, steered from desktop, mobile, web, CLI, Slack or Linear. It runs on the vendor’s infrastructure, not on your servers.
Sources: boxes.dev: Cloud computers for Claude and Codex · boxes.dev: How boxes.dev works · boxes.dev: Connect Codex and Claude · boxes.dev: Your data and credentials| Feature | devctrlSelf-hosted control plane | boxes.devHosted cloud devboxes for coding agents |
|---|---|---|
| Software pricing | Free self-hosted software | Free trial with 20 box-hours; Starter $19, Pro $99 and Max $249 per user per month with included box-hours; a team plan on custom volume |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesAgents keep working in the cloud after you close your laptop |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | LimitedExecution runs on vendor infrastructure; BYO cloud / VPC is listed for the team plan |
| Run the control plane on your server | YesRun controller and browser gateway on your server | Not documentedA self-hosted control plane is not established; BYO cloud / VPC is listed for the team plan only |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | YesA web app is listed alongside the desktop, mobile and CLI clients |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | One cloud Linux machine per devbox, created from a Template box snapshot |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Machines are full Linux environments; uploaded files and credentials are available to the agents, so the vendor advises scoped development credentials |
All features
| Feature | devctrlSelf-hosted control plane | boxes.devHosted cloud devboxes for coding agents |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Free trial with 20 box-hours; Starter $19, Pro $99 and Max $249 per user per month with included box-hours; a team plan on custom volume |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Runs your own Codex or Claude subscription or API key; boxes.dev does not resell tokens |
| Execution compute costs | Your local/remote compute bill | Vendor machines billed per awake box-hour, extra usage $0.40 per box-hour; sleeping devboxes are free |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | Desktop, iOS, Android and web apps; dvb CLI for macOS / Linux |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | Hosted service running the native Codex and Claude harnesses on managed Linux machines |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Cloud devbox per agent, steered from desktop, phone, terminal, Slack or Linear; not a controller for your own servers |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Vendor-hosted machines; the team plan lists dedicated resources and BYO cloud / VPC |
| Run agents on this computer | YesStart sessions on this computer; Docker required | Not documentedAgents run on cloud devboxes; the vendor states that nothing runs on your laptop |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | YesEach agent gets a vendor-managed cloud Linux machine, 4 vCPU / 8 GiB / 100 GB by default |
| Run the control plane on your server | YesRun controller and browser gateway on your server | Not documentedA self-hosted control plane is not established; BYO cloud / VPC is listed for the team plan only |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | One cloud Linux machine per devbox, created from a Template box snapshot |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | YesEach task gets its own devbox with separate filesystem, processes, ports and database state |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesAgents keep working in the cloud after you close your laptop |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | LimitedExecution runs on vendor infrastructure; BYO cloud / VPC is listed for the team plan |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | YesRead threads, approve agent requests and review diffs from desktop or mobile |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | YesA web app is listed alongside the desktop, mobile and CLI clients |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | LimitedIdle devboxes sleep with files and memory kept and wake later; nothing runs while asleep |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | YesSleeping keeps a machine’s files and memory; Template boxes are saved as snapshots |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | YesA Template box snapshot is the starting environment for every new devbox |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | LimitedEnvironment files and selected keys are uploaded to the machine and are available to every tool there |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | LimitedEach devbox has dedicated vCPU and memory; size is configurable per account |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | YesSetup agents port your local dev environment, or an agent gets a GitHub clone running, before the snapshot |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Continue on another subscription | YesEligible account + saved context | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | LimitedChoose model, reasoning effort and speed per thread manually; no delegated-task model advisor is established |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | LimitedPer-box permissions let agents create devboxes, organize work or update project settings |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | Not documentedRemote MCP servers can be added per project or for all projects; per-tool allow or deny rules are not established |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | YesStart and continue agents from Slack and Linear; results report back in the thread or issue |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | LimitedTeammates can follow a live transcript, diffs and screenshots; shared control is not established |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | LimitedSSO, SCIM, RBAC and audit logs are listed for the team plan |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | LimitedSpend and concurrency limits are listed for the team plan |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | YesA fresh cloud machine per body of work, started from the snapshot in seconds |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | Equivalent workflow is not established by the reviewed primary sources |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | LimitedAudit logs are listed for the team plan |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | Equivalent workflow is not established by the reviewed primary sources |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Machines are full Linux environments; uploaded files and credentials are available to the agents, so the vendor advises scoped development credentials |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | LimitedEach devbox is a separate cloud Linux machine; the isolation technology is not documented |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | LimitedNetwork allowlists are listed for the team plan |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | LimitedClaude working mode or a custom Codex / Claude command policy decides when the agent asks before edits and commands |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | boxes.dev operations need per-box permissions; file edits and commands follow the agent’s own working mode |
Cloudroom comparison
Local and cloud agent workflows with a self-hostable Linux Core. Hosted sandboxes and own-machine Core have different isolation and account boundaries.
Sources: Cloudroom: local/cloud harnesses, Teleport and subagents · Cloudroom Core: self-hosted Linux runtime and account boundary · Cloudroom: durable prompt queues and native conversation recovery · Cloudroom: free local/cloud tier and paid cloud plans| Feature | devctrlSelf-hosted control plane | CloudroomLocal/cloud agents + self-hosted Core |
|---|---|---|
| Software pricing | Free self-hosted software | Free local use with a bounded cloud tier; paid cloud plans |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesClient disconnect does not stop Core execution; native conversation recovery is separate from interrupted external effects |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesCore runs on your Linux machine without a Cloudroom account |
| Run the control plane on your server | YesRun controller and browser gateway on your server | YesSelf-hostable Core HTTP service and local history |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedNot documented in reviewed sources |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedNot documented in reviewed sources |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Hosted product advertises per-agent sandboxes/VMs; self-hosted Core uses a separate unprivileged Linux user, not an automatic per-agent VM |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedNot documented in reviewed sources |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Not documented in reviewed sources |
All features
| Feature | devctrlSelf-hosted control plane | CloudroomLocal/cloud agents + self-hosted Core |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Free local use with a bounded cloud tier; paid cloud plans |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Not documented in reviewed sources |
| Execution compute costs | Your local/remote compute bill | Not documented in reviewed sources |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | Not documented in reviewed sources |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | Apache-2.0 Rust Core; desktop app is separate |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Not documented in reviewed sources |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Local desktop, hosted cloud sandboxes, or your self-hosted Core |
| Run agents on this computer | YesStart sessions on this computer; Docker required | YesDesktop supports local agents; Core can run on your own Linux host |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | YesHosted cloud sandboxes with plan limits |
| Run the control plane on your server | YesRun controller and browser gateway on your server | YesSelf-hostable Core HTTP service and local history |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Hosted product advertises per-agent sandboxes/VMs; self-hosted Core uses a separate unprivileged Linux user, not an automatic per-agent VM |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | YesSeparate agent branches and sandbox workflows |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesClient disconnect does not stop Core execution; native conversation recovery is separate from interrupted external effects |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | Not documentedNot documented in reviewed sources |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesCore runs on your Linux machine without a Cloudroom account |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | Not documentedNot documented in reviewed sources |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedNot documented in reviewed sources |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedNot documented in reviewed sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedNot documented in reviewed sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedNot documented in reviewed sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedNot documented in reviewed sources |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedNot documented in reviewed sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedNot documented in reviewed sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedNot documented in reviewed sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedNot documented in reviewed sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedA separate frozen-session view with retained terminal snapshots and this S3 recovery workflow is not established by the reviewed sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedAn equivalent tree of agent-reported tasks and hook-recorded native steps for each session is not established by reviewed sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedAn owner-scoped inbox in which separate agent sessions exchange messages through a managed MCP tool is not established by reviewed sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedNot documented in reviewed sources |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documentedNot documented in reviewed sources |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Not documentedNot documented in reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedNot documented in reviewed sources |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | Not documentedNot documented in reviewed sources |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | Not documentedNot documented in reviewed sources |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | Not documentedNot documented in reviewed sources |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | Not documentedNot documented in reviewed sources |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | Not documentedNot documented in reviewed sources |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedNot documented in reviewed sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedNot documented in reviewed sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | Not documentedNot documented in reviewed sources |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | Not documentedNot documented in reviewed sources |
| Continue on another subscription | YesEligible account + saved context | Not documentedNot documented in reviewed sources |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | LimitedTeleport and multi-model subagent workflows documented; semantics depend on harness |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | Not documentedNot documented in reviewed sources |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedNot documented in reviewed sources |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedNot documented in reviewed sources |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedNot documented in reviewed sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedThis comparison has not established equivalent per-account native model catalogs and default selection from the cited sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedShared per-user account display ordering across clients is not established by the cited sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | YesAuthenticated Core HTTP API and child-agent tools |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | Not documentedNot documented in reviewed sources |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedNot documented in reviewed sources |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | Not documentedNot documented in reviewed sources |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | Not documentedNot documented in reviewed sources |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | Not documentedNot documented in reviewed sources |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedNot documented in reviewed sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | LimitedMulti-model subagents and child lifecycle documented; exact vendor/account policy differs |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | Not documentedNot documented in reviewed sources |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | Not documentedNot documented in reviewed sources |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedNot documented in reviewed sources |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | Not documentedNot documented in reviewed sources |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | Not documentedNot documented in reviewed sources |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documentedNot documented in reviewed sources |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | Not documentedNot documented in reviewed sources |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | Not documentedNot documented in reviewed sources |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | Not documentedNot documented in reviewed sources |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Not documentedNot documented in reviewed sources |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Not documentedNot documented in reviewed sources |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Not documentedNot documented in reviewed sources |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | Not documentedNot documented in reviewed sources |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedAn integrated per-container journal covering these tool classes and verified installation outcomes is not established by the cited sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedDeletion triggered by all coding sessions exiting is not established by reviewed sources; timeouts and inactivity policies are different |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | LimitedNative conversation resume and saved queues; interrupted dispatched work is unknown and is not blindly replayed |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | Not documented in reviewed sources |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | Not documentedNot documented in reviewed sources |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedNot documented in reviewed sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedNot documented in reviewed sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | Not documentedNot documented in reviewed sources |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedNot documented in reviewed sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | Not documented in reviewed sources |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Not documented in reviewed sources |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | Not documentedNot documented in reviewed sources |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | Not documentedNot documented in reviewed sources |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedNot documented in reviewed sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | Not documentedNot documented in reviewed sources |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | Not documented in reviewed sources |
cmux comparison
A native terminal with SSH workspaces, durable Cloud machines and CodeRouter account failover. Cloud is currently in early access.
Sources: cmux overview · cmux SSH · cmux Cloud · CodeRouter| Feature | devctrlSelf-hosted control plane | cmuxTerminal + optional Cloud |
|---|---|---|
| Software pricing | Free self-hosted software | Free terminal; paid Cloud plans from $50/month; Enterprise custom |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesSessions keep running after client disconnect; execution host must stay online |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesConnect your existing execution hosts over SSH |
| Run the control plane on your server | YesRun controller and browser gateway on your server | Not documentedSSH daemon runs on your host; Enterprise self-hosted compute does not establish a self-hosted Cloud control plane |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedDesktop / iOS; web console not documented |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | LimitedCloud: built-in vm push/pull. SSH: drag-and-drop upload; downloads use manual scp |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Local/SSH host processes; Cloud microVMs with separate Linux kernels |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesPersonal/team accounts; Enterprise SSO/SAML |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Manual setupRun Dev Containers tooling/configuration yourself; installed Docker is not a managed devcontainer workflow |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Cloud microVM boundary, machine egress policy and external model credentials |
All features
| Feature | devctrlSelf-hosted control plane | cmuxTerminal + optional Cloud |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Free terminal; paid Cloud plans from $50/month; Enterprise custom |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Your provider subscriptions/API bill |
| Execution compute costs | Your local/remote compute bill | Paid Cloud plans include pooled VM capacity; no metered overage |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | Desktop: macOS; iOS companion; TUI: macOS / Windows / Linux |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | Swift/AppKit client on Ghostty; Go remote relay |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Terminal workspaces plus optional Cloud machines and CodeRouter |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Local/SSH + vendor Cloud; Enterprise self-hosted/air-gapped execution |
| Run agents on this computer | YesStart sessions on this computer; Docker required | YesStart sessions on this computer |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | YesPersistent Linux Cloud microVMs; early-access rollout |
| Run the control plane on your server | YesRun controller and browser gateway on your server | Not documentedSSH daemon runs on your host; Enterprise self-hosted compute does not establish a self-hosted Cloud control plane |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Local/SSH host processes; Cloud microVMs with separate Linux kernels |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | Manual setupWorkspaces on a machine; branch/folder setup is task-owned |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesSessions keep running after client disconnect; execution host must stay online |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | Not documentedSSH workspace tabs and a Cloud machine tree are documented; one combined agent-session list is not |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesConnect your existing execution hosts over SSH |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | YesShows working / waiting-for-input states for supported agent sessions |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedDesktop / iOS; web console not documented |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedNot documented in reviewed sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedNot documented in reviewed sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedNot documented in reviewed sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedNot documented in reviewed sources |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedNot documented in reviewed sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedNot documented in reviewed sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedNot documented in reviewed sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedNot documented in reviewed sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedNative session resume is documented; automatic compaction after completed work is not established |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedA separate frozen-session view with retained terminal snapshots and this S3 recovery workflow is not established by the reviewed sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedAn equivalent tree of agent-reported tasks and hook-recorded native steps for each session is not established by reviewed sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedAn owner-scoped inbox in which separate agent sessions exchange messages through a managed MCP tool is not established by reviewed sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | LimitedCloud: built-in vm push/pull. SSH: drag-and-drop upload; downloads use manual scp |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documentedWatched uploads; two-way sync not documented |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Manual setupMount local/remote storage on the execution host yourself; cmux does not document managed folder grants |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Manual setupRun Dev Containers tooling/configuration yourself; installed Docker is not a managed devcontainer workflow |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | YesCloud machine disk retained through pause/resume; delete is destructive |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | YesMachine checkpoints/forks where provider supports them |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | YesSecret file uploads; CodeRouter credentials kept outside VM |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | LimitedPublish Cloud ports with personal/team access; a per-project service-grant catalog is not documented |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | YesVM sizes and plan-wide active CPU/RAM pool limits |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedNot documented in reviewed sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedNot documented in reviewed sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | YesCloud Ubuntu image with Docker, tools and agent CLIs installed |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | YesShows remaining provider account quota; configure CodeRouter |
| Continue on another subscription | YesEligible account + saved context | YesConfigured CodeRouter switches eligible accounts after quota errors |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | Not documentedNative resume; cross-agent transfer not documented |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | Not documentedAccount routing; task-model advice not documented |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedAgent teams; advice hook not documented |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedUsage views; project budgets not documented |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedNot documented in reviewed sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedThis comparison has not established equivalent per-account native model catalogs and default selection from the cited sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedShared per-user account display ordering across clients is not established by the cited sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | YesAgents create/manage workspaces and send/read terminal input through CLI/socket tools |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | Not documentedMachine network policy; project grants not documented |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedAgent approval hooks; project-grant flow not documented |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | Not documentedMachine network rules do not establish MCP tool/method filtering |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | Not documented |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | Not documented |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedNot documented in reviewed sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documented |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | Not documentedTeam machines are supported; live session-sharing roles are not documented |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | LimitedPersonal Cloud VMs can be separate; each user configures the repository. Team machines share runtime access |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesPersonal/team accounts; Enterprise SSO/SAML |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | Not documentedPersonal/team resource access documented; granular role model unspecified |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | LimitedPersonal Cloud VMs are private; team-owned machines share an environment across members |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documentedMachine sizes; aggregate per-user budgets not documented |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | YesCreate a Cloud microVM from the app; paid plan and early-access rollout required |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | Manual setupDocker is installed on Cloud machines; you create and manage agent containers yourself |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | LimitedChoose an existing/new Cloud machine; workspaces on one machine share its OS |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | LimitedCloud machine updater manages installed agent versions; separate workspace upgrade is not documented |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Manual setupMachine deletion / custom artifact cleanup |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Manual setupInstall and update tools yourself; no managed context-tools workflow documented |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | Not documentedNative app + remote daemon |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedAn integrated per-container journal covering these tool classes and verified installation outcomes is not established by the cited sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedDeletion triggered by all coding sessions exiting is not established by reviewed sources; timeouts and inactivity policies are different |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | YesRestores layout/native agent resume; Cloud daemon restart stops old processes |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | Remote daemon owns PTYs; restart recovery is layout/native agent resume |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | YesEnterprise session/network/browser audit logs |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedNot documented in reviewed sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedNot documented in reviewed sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | YesMachine CPU/memory/disk stats and pooled capacity |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedNot documented in reviewed sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | CodeRouter token usage and API-equivalent value by machine/workspace |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Cloud microVM boundary, machine egress policy and external model credentials |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | LimitedCloud microVMs provide a boundary; local/SSH agents use the permissions of their host |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | YesMachine full Internet/allowlist/no Internet modes |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedNot documented in reviewed sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | LimitedCloud VM filesystem boundary; agents can gain root inside the VM. Local/SSH restrictions are your setup |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | Rules and machine network enforcement are separate mechanisms |
Coder comparison
Self-hosted development infrastructure with Terraform workspaces and native Coder Agents. Its agent loop runs in the control plane; external Claude and Codex CLIs can also run in workspaces.
Sources: Coder platform · Coder Agents · Coder pricing| Feature | devctrlSelf-hosted control plane | CoderInfrastructure + own agent runtime |
|---|---|---|
| Software pricing | Free self-hosted software | Free Community; paid Premium / AI Premium plans |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesBrowser disconnect leaves workspace terminals/native agents running; workspace must remain started |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesYour infrastructure; Terraform defines workspaces |
| Run the control plane on your server | YesRun controller and browser gateway on your server | YesControl plane + native agent loop on your infrastructure |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | YesWeb terminal, IDEs + native Coder Agents chat |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | YesTemplate admin installs File Browser for upload/download; native agent chat also provides downloadable attachments |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Terraform selects container, pod or VM boundaries |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesNamed users, SSO, groups and workspace ownership |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Yesdevcontainer.json via Dev Containers CLI + Docker |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Configured workspace isolation; Premium governance |
All features
| Feature | devctrlSelf-hosted control plane | CoderInfrastructure + own agent runtime |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Free Community; paid Premium / AI Premium plans |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Your configured LLM provider / optional BYOK |
| Execution compute costs | Your local/remote compute bill | Your local/remote compute bill |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | Desktop/mobile web browsers; CLI and optional Coder Desktop |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | Open-source Go server + TypeScript web client |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Development infrastructure + native Coder Agents |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Self-hosted control plane + your workspace infrastructure |
| Run agents on this computer | YesStart sessions on this computer; Docker required | YesStart sessions on this computer; Docker quickstart available |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | Not documentedCoder provisions your infrastructure; Coder-operated execution compute is not established here |
| Run the control plane on your server | YesRun controller and browser gateway on your server | YesControl plane + native agent loop on your infrastructure |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Terraform selects container, pod or VM boundaries |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | Manual setupSeparate compute and security limits in templates |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesBrowser disconnect leaves workspace terminals/native agents running; workspace must remain started |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | LimitedDashboard lists native Coder Agent chats across workspaces; a combined external-CLI session list is not documented |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesYour infrastructure; Terraform defines workspaces |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | LimitedWorking/waiting states for native Coder Agents; external Claude/Codex CLI states not verified |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | YesWeb terminal, IDEs + native Coder Agents chat |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedNot documented in reviewed sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedNot documented in reviewed sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedNot documented in reviewed sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedNot documented in reviewed sources |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedNot documented in reviewed sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedNot documented in reviewed sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedNot documented in reviewed sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedNot documented in reviewed sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedNative Coder Agents compact at token thresholds or on demand; completed-task automation for Claude/Codex CLIs is not established |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedA separate frozen-session view with retained terminal snapshots and this S3 recovery workflow is not established by the reviewed sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedAn equivalent tree of agent-reported tasks and hook-recorded native steps for each session is not established by reviewed sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedAn owner-scoped inbox in which separate agent sessions exchange messages through a managed MCP tool is not established by reviewed sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | YesTemplate admin installs File Browser for upload/download; native agent chat also provides downloadable attachments |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | YesCoder Desktop bidirectional file synchronization |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Manual setupConfigure persistent volumes / mounts in Terraform |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Yesdevcontainer.json via Dev Containers CLI + Docker |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | Manual setupPersistent volumes separated from stopped compute |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | YesTerraform templates; approved infrastructure / tools |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | LimitedNative model keys stay in the controller; workspace-secret delivery requires configured templates/integrations |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | YesWorkspace app / port sharing; template infrastructure |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | Manual setupCPU / memory / security limits set by infrastructure |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedNot documented in reviewed sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedNot documented in reviewed sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | YesTerraform templates + native agent workspace provisioning |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | Not documentedAI spend tracked; CLI subscription quota unverified |
| Continue on another subscription | YesEligible account + saved context | Not documentedOptional BYOK; CLI subscription failover unverified |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | Not documentedMultiple models; Claude↔Codex CLI transfer unverified |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | LimitedNative Coder subagents choose from enabled models; no external Claude/Codex advisor integration is documented |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedNative model selection; external CLI hook unverified |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedPremium user/group AI budgets; project routing unverified |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedNot documented in reviewed sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedThis comparison has not established equivalent per-account native model catalogs and default selection from the cited sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedShared per-user account display ordering across clients is not established by the cited sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | YesNative workspace/orchestration tools + skills / MCP |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | YesAdmin-approved org MCP; template-scoped infrastructure |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedCentral MCP policy; per-project request flow unverified |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | YesAdministrators configure external MCP tool allow/deny lists for native Coder Agents |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | YesTemplate permissions + enforced administrator policies |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | Not documentedWeb/API agent chat; messenger approval flow unverified |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedNot documented in reviewed sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documentedMCP allowlists are documented; per-session call-rate caps are not |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | LimitedNative agent chats: view only with live updates; joint prompting is not documented |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | YesUsers create owned workspaces from a shared repository/template; explicit workspace sharing is separate |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesNamed users, SSO, groups and workspace ownership |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | YesBuilt-in roles; Premium custom roles / multi-org RBAC |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | YesOwned workspaces; native agents cannot cross owners |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | YesPremium per-user resource-credit quotas |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | YesTerraform provisions Docker / Kubernetes / VM workspaces |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | YesManaged Docker templates + Dev Containers integration |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | YesTemplates define container / pod / VM layout |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Manual setupMaintain template images; native harness in control plane |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Manual setupWorkspace lifecycle cleanup; session artifact manifests not documented |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Not documentedNot documented in reviewed sources |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | LimitedCoder ships one server binary; production also requires PostgreSQL |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedAn integrated per-container journal covering these tool classes and verified installation outcomes is not established by the cited sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedDeletion triggered by all coding sessions exiting is not established by reviewed sources; timeouts and inactivity policies are different |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | Manual setupStored chats + configured persistent disks; restart compute |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | Chats survive rebuilds; terminal persistence while running |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | YesPremium audit logging + HTTP / AI usage governance |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedNot documented in reviewed sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedNot documented in reviewed sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | YesWorkspace metrics + Premium configured resource credits |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedNot documented in reviewed sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | Premium AI spend export by user, group, model, provider |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Configured workspace isolation; Premium governance |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | YesPremium Agent Firewall; configured containers / VMs |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | YesPremium Agent Firewall domain / HTTP allowlists |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedNot documented in reviewed sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | Manual setupRuntime filesystem boundary follows template isolation |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | Admin policy enforced server-side; instructions add context |
Conductor comparison
Parallel workspaces on your Mac or managed Cloud sandboxes, with team collaboration and API access. Cloud agents can keep working after the app closes.
Sources: Conductor model · Conductor Cloud · Conductor API| Feature | devctrlSelf-hosted control plane | ConductorLocal / managed Cloud workspace |
|---|---|---|
| Software pricing | Free self-hosted software | Free local; Pro $50/month; Teams $60/user/month; Enterprise custom |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | LimitedCloud agents continue when the app closes; local Mac sessions end when the app quits |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | NoCloud self-hosting not yet available |
| Run the control plane on your server | YesRun controller and browser gateway on your server | NoLocal app self-managed; self-hosted Cloud control plane not available |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedDesktop / iOS / API; browser console not documented |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedCloud/local sync; direct transfers not documented |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Local processes/worktrees; isolated microVM-backed Cloud sandboxes |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesTeam/organization accounts and user roles |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedCloud Computer setup documented; devcontainer config lifecycle not covered |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Cloud sandbox boundary; local user permissions and native tool approvals |
All features
| Feature | devctrlSelf-hosted control plane | ConductorLocal / managed Cloud workspace |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Free local; Pro $50/month; Teams $60/user/month; Enterprise custom |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Your provider subscriptions/API bill |
| Execution compute costs | Your local/remote compute bill | Local: your compute bill. Cloud: capacity included in paid plans |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | macOS desktop, iOS companion and cloud API |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | Desktop/cloud implementation stack not specified in reviewed primary docs |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Agent workspace/review app with managed Cloud sandboxes |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Local Mac or vendor-managed Cloud; Cloud self-host not yet available |
| Run agents on this computer | YesStart sessions on this computer; Docker required | YesStart sessions on this computer; macOS required |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | YesManaged Vercel sandboxes/Cloud microVMs |
| Run the control plane on your server | YesRun controller and browser gateway on your server | NoLocal app self-managed; self-hosted Cloud control plane not available |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Local processes/worktrees; isolated microVM-backed Cloud sandboxes |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | YesWorkspace branch/files and separate Cloud sandbox per workspace |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | LimitedCloud agents continue when the app closes; local Mac sessions end when the app quits |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | Not documentedMac + managed Cloud; SSH fleet not documented |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | NoCloud self-hosting not yet available |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | YesShows working / waiting-for-input states for supported agent sessions |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedDesktop / iOS / API; browser console not documented |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedNot documented in reviewed sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedNot documented in reviewed sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedNot documented in reviewed sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedNot documented in reviewed sources |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedNot documented in reviewed sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedNot documented in reviewed sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedNot documented in reviewed sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedNot documented in reviewed sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedAutomatic server-side compaction after completed work is not established by reviewed sources |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedA separate frozen-session view with retained terminal snapshots and this S3 recovery workflow is not established by the reviewed sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedAn equivalent tree of agent-reported tasks and hook-recorded native steps for each session is not established by reviewed sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedAn owner-scoped inbox in which separate agent sessions exchange messages through a managed MCP tool is not established by reviewed sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedCloud/local sync; direct transfers not documented |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | YesCloud / local automated file sync |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Manual setupConfigure local directories or tool integrations yourself; general remote-folder attachment is not documented |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedCloud Computer setup documented; devcontainer config lifecycle not covered |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | LimitedLocal files persist; Cloud disk is ephemeral and requires supported sync/retention workflows |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | YesCloud Computer preinstalled repo/dependencies and setup settings |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | YesEnvironment variables/secrets configured for workspace sandbox |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | Manual setupConfigured MCP/tools and environment; service-grant broker not documented |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | YesDefined Cloud sandbox CPU/RAM/disk; aggregate per-user budgets unspecified |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedNot documented in reviewed sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedNot documented in reviewed sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | YesCloud repos/dependencies preinstalled; local repository setup scripts |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | YesCodex account limits + reset times |
| Continue on another subscription | YesEligible account + saved context | Not documentedBring subscriptions; continuation flow not documented |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | Not documentedMultiple agents; context transfer not documented |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | Not documentedModel picker; task advice not documented |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documented |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documented |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedNot documented in reviewed sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedThis comparison has not established equivalent per-account native model catalogs and default selection from the cited sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedShared per-user account display ordering across clients is not established by the cited sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | YesAgents create/manage Cloud workspaces through the Cloud API/MCP |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | Not documentedWorkspace MCP / secrets; grant flow not documented |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedNative tool approvals; project grants not documented |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | Not documentedNative agent approvals do not establish a managed MCP tool allowlist |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | Not documented |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | Manual setupBuild a bot with Cloud API |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedNot documented in reviewed sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documented |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | YesCloud Multiplayer: view and prompt together; paid plan required |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | YesTeam members create workspaces for the same repository; Cloud gives each workspace a sandbox |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesTeam/organization accounts and user roles |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | YesTeam role administration; Enterprise identity/security controls |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | Not documentedWorkspace sandbox isolation; private per-user tenant policy not specified |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documentedPlan / sandbox limits; per-user aggregate budgets not documented |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | YesCreating a Cloud workspace also creates its isolated sandbox |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | Manual setupConfigure a container/custom CLI wrapper yourself; built-in Cloud uses Vercel sandboxes |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | YesStart chats in an existing workspace or create a separate Cloud sandbox |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Manual setupBundled CLI / custom executable |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Manual setupWorkspace archive scripts |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Manual setupInstall and update tools yourself; no managed context-tools workflow documented |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | Not documentedDesktop app + managed Cloud |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedAn integrated per-container journal covering these tool classes and verified installation outcomes is not established by the cited sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedDeletion triggered by all coding sessions exiting is not established by reviewed sources; timeouts and inactivity policies are different |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | Not documentedCloud agent persistence documented; whole sandbox/server recovery not specified |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | Cloud runs independently of app; closing local app ends local sessions |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | Not documentedEnterprise security controls; detailed operation audit not covered by reviewed docs |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedNot documented in reviewed sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedNot documented in reviewed sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | Not documentedPlan/sandbox capacity described; detailed live accounting unspecified |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedNot documented in reviewed sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | Quota views; per-workspace invoice attribution not documented |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Cloud sandbox boundary; local user permissions and native tool approvals |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | LimitedCloud sandboxes isolate execution; local agents run with the Mac user’s permissions |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | Not documentedNetwork behavior documented; workspace egress policy not specified |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedNot documented in reviewed sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | LimitedCloud sandboxes separate filesystems; local agents retain the Mac user’s host file access |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | Repository guidance differs from Cloud/OS and tool approval boundaries |
Daytona comparison
Sandbox infrastructure for agent applications: SDKs, API, containers and VMs. Hosted compute or Enterprise BYOC; documented BYOC keeps Daytona’s control plane. Current core development is private.
Sources: Daytona platform · Daytona BYOC · Daytona source| Feature | devctrlSelf-hosted control plane | DaytonaSandbox infrastructure API |
|---|---|---|
| Software pricing | Free self-hosted software | Usage-based cloud compute; custom Enterprise / BYOC |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | LimitedSandbox processes survive client disconnect; configure inactivity/TTL policies or they can stop automatically |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesEnterprise BYOC runner nodes on your infrastructure |
| Run the control plane on your server | YesRun controller and browser gateway on your server | Not documentedBYOC keeps Daytona control plane; old public core frozen |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | YesDashboard + authenticated sandbox web terminal |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | YesSDK/API streaming file uploads and downloads |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Isolated containers or full Linux / Windows VMs |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesUsers + personal/collaborative orgs; Enterprise SSO |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedDocker/OCI snapshots; devcontainer.json flow unverified |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Sandbox isolation; organization is runtime trust boundary |
All features
| Feature | devctrlSelf-hosted control plane | DaytonaSandbox infrastructure API |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Usage-based cloud compute; custom Enterprise / BYOC |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Your provider subscriptions/API bill |
| Execution compute costs | Your local/remote compute bill | Reserved CPU / RAM / disk; cloud billing per second |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | Desktop/mobile web browsers; SDK/CLI |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | Current core private; public SDKs in multiple languages |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Sandbox infrastructure/API for your agent application |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Daytona cloud; Enterprise BYOC compute runners |
| Run agents on this computer | YesStart sessions on this computer; Docker required | Manual setupInstall/configure Enterprise BYOC runners yourself; a standalone local agent-workspace workflow is not documented |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | YesHosted sandbox compute; optional Enterprise BYOC |
| Run the control plane on your server | YesRun controller and browser gateway on your server | Not documentedBYOC keeps Daytona control plane; old public core frozen |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Isolated containers or full Linux / Windows VMs |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | Manual setupCreate separate sandboxes and check out repositories with your application; no task-worktree manager is documented |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | LimitedSandbox processes survive client disconnect; configure inactivity/TTL policies or they can stop automatically |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | Not documentedSandbox dashboard; native CLI session fleet unverified |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesEnterprise BYOC runner nodes on your infrastructure |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | Not documentedSandbox states; agent working/waiting UI unverified |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | YesDashboard + authenticated sandbox web terminal |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedNot documented in reviewed sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedNot documented in reviewed sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedNot documented in reviewed sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedNot documented in reviewed sources |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedNot documented in reviewed sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedNot documented in reviewed sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedNot documented in reviewed sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedNot documented in reviewed sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedSandbox infrastructure is documented; automatic compaction of completed coding sessions is not established |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedA separate frozen-session view with retained terminal snapshots and this S3 recovery workflow is not established by the reviewed sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedAn equivalent tree of agent-reported tasks and hook-recorded native steps for each session is not established by reviewed sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedAn owner-scoped inbox in which separate agent sessions exchange messages through a managed MCP tool is not established by reviewed sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | YesSDK/API streaming file uploads and downloads |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documentedFile API; automatic folder mirroring unverified |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Manual setupVolumes; external storage through configured FUSE mounts |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedDocker/OCI snapshots; devcontainer.json flow unverified |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | YesContainer / VM disks + volumes; GPU disks ephemeral |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | YesReusable snapshots + images; container declarative builds |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | YesHTTPS header substitution; configure allowed hosts |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | LimitedLink sandboxes and share volumes; collaborative-organization members share runtime access |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | YesSandbox vCPU / RAM / disk reservations enforced |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedNot documented in reviewed sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedNot documented in reviewed sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | YesSnapshots / OCI images; install packages inside sandbox |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | Not documentedCompute quotas; agent-provider account quota unverified |
| Continue on another subscription | YesEligible account + saved context | Not documentedNative provider-subscription continuation unverified |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | Not documentedNative Claude↔Codex conversation handoff unverified |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | Manual setupSelect models in the agent application using the SDK |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedSandbox API; native CLI routing hook unverified |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedSandbox costs; project model-routing budgets unverified |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedNot documented in reviewed sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedThis comparison has not established equivalent per-account native model catalogs and default selection from the cited sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedShared per-user account display ordering across clients is not established by the cited sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | YesMCP + SDK file/process/Git/computer-use tools |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | LimitedPer-sandbox secrets/HTTPS allowlists; your application assigns project scope |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedResource RBAC; project approval/delivery flow unverified |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | Not documentedMCP sandbox tools exist; per-tool authorization by agent session is not documented |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | Manual setupConfigure sandbox networks, secrets and org roles |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | Manual setupBuild chat integration around sandbox SDK/API |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedNot documented in reviewed sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documentedPlatform API quotas are not repository/session MCP call-rate limits |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | Not documentedOrganization members share sandbox access; live agent-conversation sharing is not documented |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | Manual setupYour application clones the repository into separate user sandboxes; a team organization shares runtime access by default |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesUsers + personal/collaborative orgs; Enterprise SSO |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | YesResource roles; assignments do not restrict runtime access |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | LimitedPersonal organizations are private; members of a collaborative organization share sandbox runtime access |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documentedQuotas apply to an organization; a shared team pool is not an aggregate per-member limit |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | YesManaged sandbox lifecycle; containers / VMs / GPUs |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | YesOCI/Docker-compatible container images + snapshots |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | Manual setupYour agent application decides whether to reuse a sandbox or create another through the SDK |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Manual setupMaintain snapshots or install packages in the sandbox |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Manual setupSandbox auto-delete; agent artifact manifests are your application |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Not documentedNot documented in reviewed sources |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | Not documentedCLI client; single controller binary unverified |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedAn integrated per-container journal covering these tool classes and verified installation outcomes is not established by the cited sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedDeletion triggered by all coding sessions exiting is not established by reviewed sources; timeouts and inactivity policies are different |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | Not documentedDisk/VM pause persistence; runner reboot recovery unverified |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | VM pause preserves RAM; container stop clears processes |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | YesEnterprise organization/user/sandbox lifecycle audit logs |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedNot documented in reviewed sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedNot documented in reviewed sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | LimitedReports sandbox allocations and billed resource usage; live CPU/RAM utilization is not established here |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedNot documented in reviewed sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | Organization wallet + per-sandbox cost breakdown |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Sandbox isolation; organization is runtime trust boundary |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | YesContainer namespaces or hardware VM boundary |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | YesSandbox firewall: block all / CIDR / domain allowlists |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedNot documented in reviewed sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | YesSeparate filesystems + volume subpath isolation |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | Sandbox policy is configured through platform controls |
Emdash comparison
Open-source desktop workspace for native coding agents, Git worktrees and review, with SSH execution and optional tmux persistence.
Sources: Emdash overview · Emdash SSH · Emdash persistence| Feature | devctrlSelf-hosted control plane | EmdashDesktop agent workspace |
|---|---|---|
| Software pricing | Free self-hosted software | Free desktop; Cloud / Enterprise by contact |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | LimitedAgents keep running if the project’s tmux toggle is enabled; local Windows is excluded |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesConnect your existing execution hosts over SSH |
| Run the control plane on your server | YesRun controller and browser gateway on your server | Not documentedDesktop orchestration + managed SSH workspace daemon; central multi-user controller not documented |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedDesktop UI with app previews; browser session console not documented |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedRemote file reads / writes; dedicated upload or archive download workflow not documented |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Native CLI processes on local / SSH host |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedOS/SSH and GitHub accounts exist; separate application-user identities are not documented |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedProvision scripts support containers; Dev Container specification not documented |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Native agent permissions plus SSH / provisioned-host controls |
All features
| Feature | devctrlSelf-hosted control plane | EmdashDesktop agent workspace |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Free desktop; Cloud / Enterprise by contact |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Your provider subscriptions/API bill |
| Execution compute costs | Your local/remote compute bill | Your local/remote compute bill; optional Emdash Cloud terms apply separately |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | macOS / Windows / Linux |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | Electron / TypeScript; React UI; SQLite state |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Desktop workspace around native CLI agents |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Local projects, your SSH hosts; contact-based Cloud |
| Run agents on this computer | YesStart sessions on this computer; Docker required | YesStart sessions on this computer |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | YesEmdash Cloud offered by contact |
| Run the control plane on your server | YesRun controller and browser gateway on your server | Not documentedDesktop orchestration + managed SSH workspace daemon; central multi-user controller not documented |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Native CLI processes on local / SSH host |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | YesWorktree branches; provisioned task environments when enabled |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | LimitedAgents keep running if the project’s tmux toggle is enabled; local Windows is excluded |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | Not documentedLocal/SSH projects are listed; one combined multi-host agent-session list is not documented |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesConnect your existing execution hosts over SSH |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | YesShows working / waiting-for-input states for supported agent sessions |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedDesktop UI with app previews; browser session console not documented |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedNot documented in reviewed sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedNot documented in reviewed sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedNot documented in reviewed sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedNot documented in reviewed sources |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedNot documented in reviewed sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedNot documented in reviewed sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedNot documented in reviewed sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedNot documented in reviewed sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedAutomatic server-side compaction after completed work is not established by reviewed sources |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedA separate frozen-session view with retained terminal snapshots and this S3 recovery workflow is not established by the reviewed sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedAn equivalent tree of agent-reported tasks and hook-recorded native steps for each session is not established by reviewed sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedAn owner-scoped inbox in which separate agent sessions exchange messages through a managed MCP tool is not established by reviewed sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedRemote file reads / writes; dedicated upload or archive download workflow not documented |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documentedPreserve-pattern worktree copies; two-way folder sync not documented |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Manual setupSelected CLI permissions and host paths |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedProvision scripts support containers; Dev Container specification not documented |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | YesLocal SQLite state; local / remote project files |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | LimitedReuse setup/run/teardown scripts; prepared compute images remain your infrastructure setup |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | Manual setupPreserve .env patterns or supply provisioned-host credentials |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | Manual setupYour setup scripts and remote backend |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | Manual setupSet limits in your provisioned VM / pod / container |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedNot documented in reviewed sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedNot documented in reviewed sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | YesShell setup and lifecycle scripts; remote CLI installation |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | YesShows remaining provider account quota and reset timing |
| Continue on another subscription | YesEligible account + saved context | Not documentedNative subscriptions; quota-triggered account continuation not documented |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | Not documentedProvider resume; cross-native-agent transcript handoff not documented |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | Not documentedModel selector; task-specific recommendation not documented |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedLifecycle status hooks; model-advice hook not documented |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedQuota meters; project routing budgets not documented |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedNot documented in reviewed sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedThis comparison has not established equivalent per-account native model catalogs and default selection from the cited sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedShared per-user account display ordering across clients is not established by the cited sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | Not documentedMCP configuration sync is supported; tools for agents to manage Emdash sessions are not documented |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | Not documentedMCP credential setup; central project-resource grants not documented |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedNative agent prompts; controller-delivered project grants not documented |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | Manual setupConfigure allow/deny rules in the selected native agent; Emdash syncs MCP server configuration |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | Manual setupMaintain reusable access rules in native CLI config; shared lifecycle scripts are not grant templates |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | Not documentedIssue integrations; messenger session bots not documented |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedNot documented in reviewed sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documented |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | Not documentedTask chats are documented; live cross-user session sharing is not |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | Manual setupEach teammate opens a checkout on their host/SSH user; shared-server per-user environments need external setup |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedOS/SSH and GitHub accounts exist; separate application-user identities are not documented |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | Not documentedCloud admin controls mentioned; role permissions not specified |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | Manual setupUse separate host users / provisioned task infrastructure |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documentedCustom infrastructure limits; aggregate per-user controller budgets not documented |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | Manual setupProvision / terminate scripts when remote-task feature is enabled |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | Manual setupYour provision script can create Docker workspaces |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | YesShared SSH project or configured per-task remote workspace |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Manual setupEmdash installs its remote daemon/agent tools; per-workspace CLI image upgrades remain your setup |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Manual setupTask / tmux cleanup and your teardown script; session artifact lists not documented |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Not documentedNot documented in reviewed sources |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | Not documentedElectron app + packaged SSH workspace daemon; standalone central controller not documented |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedAn integrated per-container journal covering these tool classes and verified installation outcomes is not established by the cited sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedDeletion triggered by all coding sessions exiting is not established by reviewed sources; timeouts and inactivity policies are different |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | Not documentedtmux survives app reconnects; host-reboot recovery not documented |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | Configured tmux keeps actual agent / terminal processes alive |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | Not documentedTask diffs / CI checks; administrative audit trail not documented |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedNot documented in reviewed sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedNot documented in reviewed sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | Not documentedNative quota meters; CPU / memory / disk accounting not documented |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedNot documented in reviewed sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | Provider account usage; project compute chargeback not documented |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Native agent permissions plus SSH / provisioned-host controls |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | Manual setupSelected CLI sandbox and your provisioned infrastructure |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | Manual setupConfigure restrictions on your execution host / backend |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedNot documented in reviewed sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | Manual setupApp remote file operations stay within root; CLI / host policies govern agent access |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | Prompts / skills configure agents; host policies govern execution |
Happy comparison
Encrypted desktop, phone and browser control. Its original CLI wraps Claude and Codex; newer Desktop uses Happy Harness, with a self-hosted remote runtime.
Sources: Happy CLI · Happy remote · Happy models| Feature | devctrlSelf-hosted control plane | HappyCLI companion / own Harness |
|---|---|---|
| Software pricing | Free self-hosted software | Free self-hosted software |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesSessions keep running after client disconnect; execution host must stay online |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesSelf-hosted Happy Agent runtime |
| Run the control plane on your server | YesRun controller and browser gateway on your server | YesSelf-host Happy Agent daemon/remote runtime |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | LimitedOriginal Happy CLI has a web client; current Harness remote guide specifies desktop/phone control |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedEncrypted images + file reader; archive downloads not documented |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | OS command sandbox; configured Docker execution; task Git worktrees |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesTeam/user identities and session invites |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedDocker image configuration; devcontainer.json lifecycle not documented |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Unified OS sandbox and permissions; encrypted relay is a transport boundary |
All features
| Feature | devctrlSelf-hosted control plane | HappyCLI companion / own Harness |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Free self-hosted software |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Your provider subscriptions/API bill |
| Execution compute costs | Your local/remote compute bill | Your local/remote compute bill |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | Desktop: macOS / Windows / Linux; iOS/Android; original CLI web client |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | TypeScript agent runtime; separate desktop, CLI, mobile and relay packages |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Original native-CLI wrapper plus newer unified Happy Harness |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Local/self-hosted Happy Agent; encrypted relay/remote transport |
| Run agents on this computer | YesStart sessions on this computer; Docker required | YesStart sessions on this computer |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | Not documentedVendor compute provisioning not described; remote guide uses your server |
| Run the control plane on your server | YesRun controller and browser gateway on your server | YesSelf-host Happy Agent daemon/remote runtime |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | OS command sandbox; configured Docker execution; task Git worktrees |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | YesManaged Git worktrees and branches for parallel tasks |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesSessions keep running after client disconnect; execution host must stay online |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | Not documentedDesktop connects to local/remote runtimes; one combined cross-host session list is not documented |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesSelf-hosted Happy Agent runtime |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | YesShows working / waiting-for-input states for supported agent sessions |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | LimitedOriginal Happy CLI has a web client; current Harness remote guide specifies desktop/phone control |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedNot documented in reviewed sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedNot documented in reviewed sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedNot documented in reviewed sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedNot documented in reviewed sources |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedNot documented in reviewed sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedNot documented in reviewed sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedNot documented in reviewed sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedNot documented in reviewed sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedAutomatic server-side compaction after completed work is not established by reviewed sources |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedA separate frozen-session view with retained terminal snapshots and this S3 recovery workflow is not established by the reviewed sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedAn equivalent tree of agent-reported tasks and hook-recorded native steps for each session is not established by reviewed sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedAn owner-scoped inbox in which separate agent sessions exchange messages through a managed MCP tool is not established by reviewed sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedEncrypted images + file reader; archive downloads not documented |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documentedProject → worktree copies; remote two-way sync not documented |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Manual setupConfigured host mounts + protected paths |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedDocker image configuration; devcontainer.json lifecycle not documented |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | YesSQLite durable sessions and disk-backed project/worktree files |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | LimitedReuse project setup/copy configuration; prepared compute snapshots are not established |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | YesNamed bundles scoped to a session or project |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | YesMachine/project MCP entries and team runtime resources |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | Not documentedOS sandbox permissions are documented; managed CPU/RAM/disk caps are not |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedNot documented in reviewed sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedNot documented in reviewed sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | YesWorktree setup scripts and configuration by built-in Chief of Staff |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | YesShows remaining provider account quota |
| Continue on another subscription | YesEligible account + saved context | LimitedHappy Harness pools fail over after quota errors; native CLI conversation/account continuation is not established |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | Not documentedHarness provider switch; native CLI transfer not documented |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | Not documentedRuntime model choices; task advisor not documented |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedRuntime delegation; CLI advice hook not documented |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedQuota views; project routing budgets not documented |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedNot documented in reviewed sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedThis comparison has not established equivalent per-account native model catalogs and default selection from the cited sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedShared per-user account display ordering across clients is not established by the cited sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | YesChief of Staff configures runtime/projects; agents also use workspace MCP/skills |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | YesProject network config + session secret bundles |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedAction approvals; project-grant delivery not documented |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | LimitedHarness reviews MCP actions; a reusable per-method allowlist is not established |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | Manual setupMachine / project TOML policies |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | Not documentedNative phone client; messenger bots not documented |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedNot documented in reviewed sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documented |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | YesHarness invitations/team runtime: view and steer the same live conversation |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | LimitedTeam members use task worktrees in shared projects; the team runtime and its model accounts remain shared |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesTeam/user identities and session invites |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | Not documentedTeam/session access exists; granular resource role matrix not documented |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | Manual setupSeparate runtimes if private tenant compute is needed; team runtime shares state |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documentedPer-user aggregate compute limits not documented |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | LimitedHarness provides an OS sandbox; Docker requires a configured local image, not a cloud-resource provisioner |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | LimitedHarness executes in a configured Docker image; you prepare/pull that image yourself |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | Not documentedWorktrees; shared/separate container choice not documented |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Manual setupServer upgrades; maintain your tool images |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Manual setupWorktree archive; custom artifact lists |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Manual setupInstall and update tools on the execution host yourself |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | YesStandalone Happy Agent server binary |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedAn integrated per-container journal covering these tool classes and verified installation outcomes is not established by the cited sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedDeletion triggered by all coding sessions exiting is not established by reviewed sources; timeouts and inactivity policies are different |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | YesRuntime resumes durable sessions, waits and subagents from disk |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | Daemon survives client close; restart reconstructs durable agent state |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | Not documentedDurable events/state documented; administrator operation-audit product unspecified |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedNot documented in reviewed sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedNot documented in reviewed sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | Not documentedRuntime process management exists; infrastructure accounting unspecified |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedNot documented in reviewed sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | Quota views; per-project cost attribution not documented |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Unified OS sandbox and permissions; encrypted relay is a transport boundary |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | YesOS sandbox for commands, with reviewed boundary crossings |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | YesMachine/project network policy and sandbox boundary controls |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedNot documented in reviewed sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | YesProtected paths and workspace write/read-only sandbox modes |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | Agent guidance differs from sandbox/permission enforcement |
Herdr comparison
A terminal server with persistent workspaces across SSH hosts. Community extensions add browser access and quota-aware routing.
Sources: Herdr model · Herdr SSH · Mobile Relay · Agent Router| Feature | devctrlSelf-hosted control plane | HerdrNative CLI terminal runtime |
|---|---|---|
| Software pricing | Free self-hosted software | Free self-hosted software |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesSessions keep running after client disconnect; execution host must stay online |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesConnect your existing execution hosts over SSH |
| Run the control plane on your server | YesRun controller and browser gateway on your server | YesRun your own Herdr server and attach clients |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | With extensionInstall Mobile Relay to send prompts and terminal input from its web app |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | LimitedMobile Relay sends attachments and previews files; archive downloads require external tools |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Default: host OS processes. Containers/sandboxes require your wrapper or prepared environment |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Manual setupSSH/OS identity; no documented application-user tenancy |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Manual setupCreate the devcontainer yourself, then start Herdr/agents inside it |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | OS/SSH access and externally selected execution environment |
All features
| Feature | devctrlSelf-hosted control plane | HerdrNative CLI terminal runtime |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Free self-hosted software |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Your provider subscriptions/API bill |
| Execution compute costs | Your local/remote compute bill | Your local/remote compute bill |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | macOS / Windows / Linux |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | Rust server/client terminal runtime |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Terminal runtime hosting existing CLI agents |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Background servers on existing local/SSH machines |
| Run agents on this computer | YesStart sessions on this computer; Docker required | YesStart sessions on this computer |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | Not documentedCloud connection offering announced; managed compute not documented |
| Run the control plane on your server | YesRun controller and browser gateway on your server | YesRun your own Herdr server and attach clients |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Default: host OS processes. Containers/sandboxes require your wrapper or prepared environment |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | Manual setupSeparate working directories/worktrees you prepare |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesSessions keep running after client disconnect; execution host must stay online |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | YesAgent list combines connected machines and shows where each agent runs |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesConnect your existing execution hosts over SSH |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | YesShows working / waiting-for-input states for supported agent sessions |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | With extensionInstall Mobile Relay to send prompts and terminal input from its web app |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedNot documented in reviewed sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedNot documented in reviewed sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedNot documented in reviewed sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedNot documented in reviewed sources |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedNot documented in reviewed sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedNot documented in reviewed sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedNot documented in reviewed sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedNot documented in reviewed sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedAutomatic server-side /compact after completed work is not established by reviewed sources |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedA separate frozen-session view with retained terminal snapshots and this S3 recovery workflow is not established by the reviewed sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedAn equivalent tree of agent-reported tasks and hook-recorded native steps for each session is not established by reviewed sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedAn owner-scoped inbox in which separate agent sessions exchange messages through a managed MCP tool is not established by reviewed sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | LimitedMobile Relay sends attachments and previews files; archive downloads require external tools |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documented |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Manual setupMount storage and set permissions on the host yourself; no integrated folder-grant workflow is documented |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Manual setupCreate the devcontainer yourself, then start Herdr/agents inside it |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | YesSession/layout data and working files on supplied hosts |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | Not documentedPrepared compute/runtime template lifecycle not documented |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | Manual setupHost agent logins/environment; narrow delivery is your setup |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | Manual setupConfigure accessible services through the host environment |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | Manual setupUse host cgroups, containers or VM limits externally |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedNot documented in reviewed sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedNot documented in reviewed sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | Manual setupPrepare repository, agent tools and credentials on host |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | With extensionShows remaining provider account quota; install Agent Router |
| Continue on another subscription | YesEligible account + saved context | Not documentedNo documented continuation flow |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | Not documentedNo documented transfer flow |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | With extensionAgent Router recommends agent/model choices for tasks |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedTask recommendations do not establish a native before-spawn hook |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documented |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedNot documented in reviewed sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedThis comparison has not established equivalent per-account native model catalogs and default selection from the cited sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedShared per-user account display ordering across clients is not established by the cited sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | YesAgents call Herdr CLI/socket management API through its skill |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | Manual setupConfigure service access and credentials in the host/agent environment yourself |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedNative agent approvals; project grants not documented |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | Not documentedManaged MCP tool/method authorization is not documented |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | Not documented |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | Not documentedMessenger session bots are not covered by reviewed sources |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedNot documented in reviewed sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documented |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | Manual setupMultiple clients can view/type using supplied OS/SSH access; application-user invitations are not documented |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | Manual setupSet up separate OS users/checkouts or containers yourself; private application-user environments are not documented |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Manual setupSSH/OS identity; no documented application-user tenancy |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | Not documentedApplication-level role/resource authorization not documented |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | Manual setupSeparate OS users/hosts or external container setup |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documented |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | Manual setupPrepare an isolated host/container/sandbox yourself, then run Herdr there |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | Manual setupCreate the container and run Herdr/agents inside it yourself |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | Manual setupUse your existing environment or prepare a separate container yourself |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Manual setupUpgrade agent CLIs on the host or in your container yourself |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Manual setupKeep an artifact list and delete its files/containers yourself |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Manual setupInstall and update command filters or search tools in the host environment yourself |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | YesDownload one Herdr binary; no separate database service |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedAn integrated per-container journal covering these tool classes and verified installation outcomes is not established by the cited sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedDeletion triggered by all coding sessions exiting is not established by reviewed sources; timeouts and inactivity policies are different |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | YesRestores layout and resumes supported agent sessions after machine restart |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | Server PTYs survive client detach; restarted agents resume saved sessions |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | Not documentedInfrastructure authorization audit trail not documented |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedNot documented in reviewed sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedNot documented in reviewed sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | Not documentedPer-workspace compute accounting not documented |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedNot documented in reviewed sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | Core cost attribution not documented; plugins are separate |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | OS/SSH access and externally selected execution environment |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | Manual setupExternal sandbox wrappers supported in agent detection |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | Manual setupHost/network configuration; runtime egress policy not documented |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedNot documented in reviewed sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | Manual setupHost permissions or externally configured sandbox |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | Agent guidance and host access; no documented managed resource broker |
Nimbalyst comparison
Open-source visual workspace around native Claude and Codex, with searchable sessions, worktrees, iOS control and shared team documents.
Sources: Nimbalyst overview · Nimbalyst mobile · Nimbalyst Teams| Feature | devctrlSelf-hosted control plane | NimbalystVisual agent workspace |
|---|---|---|
| Software pricing | Free self-hosted software | Free individual; Teams beta free, then $20/user/month; Enterprise custom |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | Not documentedSearchable / resumable history; uninterrupted running after desktop exit not documented |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | Not documentedDesktop native execution; SSH execution-server setup not documented |
| Run the control plane on your server | YesRun controller and browser gateway on your server | Not documentedEnterprise self-hosts document collaboration; a remote agent-execution control plane is not documented |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedTeams browser console edits shared docs; agent execution control is desktop / iOS |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedFile links / mobile markdown; direct file or archive transfer workflow not documented |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Native local agents; iOS controls paired desktop execution |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesTeam accounts / organization membership; paired mobile accounts |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedWorktree sessions; Dev Container execution workflow not documented |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Project trust plus native review, action / path / URL approvals |
All features
| Feature | devctrlSelf-hosted control plane | NimbalystVisual agent workspace |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Free individual; Teams beta free, then $20/user/month; Enterprise custom |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Your provider subscriptions/API bill |
| Execution compute costs | Your local/remote compute bill | Your local compute bill |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | macOS / Windows / Linux; iOS companion |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | TypeScript / Electron / React; SwiftUI iOS; SQLite (legacy PGLite) |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Visual desktop workspace around native coding agents |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Agents execute on desktop; shared docs use cloud collaboration service |
| Run agents on this computer | YesStart sessions on this computer; Docker required | YesStart sessions on this computer |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | Not documentedHosted collaboration; vendor agent compute not documented |
| Run the control plane on your server | YesRun controller and browser gateway on your server | Not documentedEnterprise self-hosts document collaboration; a remote agent-execution control plane is not documented |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Native local agents; iOS controls paired desktop execution |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | YesSeparate worktree branches / checkouts; multiple sessions can share a worktree |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | Not documentedSearchable / resumable history; uninterrupted running after desktop exit not documented |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | Not documentediOS accesses a paired desktop; one multi-host agent-session list is not documented |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | Not documentedDesktop native execution; SSH execution-server setup not documented |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | YesShows working / waiting-for-input states for supported agent sessions |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedTeams browser console edits shared docs; agent execution control is desktop / iOS |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedNot documented in reviewed sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedNot documented in reviewed sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedNot documented in reviewed sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedNot documented in reviewed sources |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedNot documented in reviewed sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedNot documented in reviewed sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedNot documented in reviewed sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedNot documented in reviewed sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedA Compact action is documented; automatic server-side compaction after completed work is not established |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedA separate frozen-session view with retained terminal snapshots and this S3 recovery workflow is not established by the reviewed sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedAn equivalent tree of agent-reported tasks and hook-recorded native steps for each session is not established by reviewed sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedAn owner-scoped inbox in which separate agent sessions exchange messages through a managed MCP tool is not established by reviewed sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedFile links / mobile markdown; direct file or archive transfer workflow not documented |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documentedLive shared docs / trackers; alpha mobile markdown; general two-way folder sync not documented |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | LimitedAdd local directories in native agent settings; remote-storage transport and read/write mounts are not documented |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedWorktree sessions; Dev Container execution workflow not documented |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | YesOrdinary local project files; SQLite session history and rolling backups |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | Manual setupPrepare local dependencies yourself; MCP/command templates do not prepare execution compute |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | YesMCP credentials stored locally; app or project server scope |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | LimitedConfigure project MCP connections; central shared-resource publication/grants are not documented |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | Not documentedLocal agent execution; host CPU / RAM caps not documented |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedNot documented in reviewed sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedNot documented in reviewed sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | Manual setupAgent configuration and worktrees; maintain local dependencies |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | YesShows remaining provider account quota; token activity reports are separate |
| Continue on another subscription | YesEligible account + saved context | Not documentedBring native subscriptions; quota-triggered account continuation not documented |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | Not documentedMultiple native providers / session branches; cross-provider transcript transfer not documented |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | Not documentedModel picker; task-specific model advice not documented |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedExperimental agent teams; native CLI model-advice hook not documented |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedProject token reports / loop iteration limits; routing budgets not documented |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedNot documented in reviewed sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedThis comparison has not established equivalent per-account native model catalogs and default selection from the cited sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedShared per-user account display ordering across clients is not established by the cited sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | YesNative workspace MCP tools, trackers, visual editors and extension tools |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | LimitedProject-scoped MCP connections/credentials; general API/database service-grant routing is not documented |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedInline project action approvals; controller-delivered resource grants not documented |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | LimitedEnable project MCP servers and native approvals; per-method broker rules are not documented |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | LimitedReuse native project approval patterns; resource-grant templates across sessions are not documented |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | Not documentedBuilt-in team chat and iOS agent messages; external messenger bots not documented |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedNot documented in reviewed sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documented |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | Not documentedSession links are documented; live output sharing/control is not established |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | LimitedTeammates use their own local checkouts/agents; shared documents do not create shared-server environments |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesTeam accounts / organization membership; paired mobile accounts |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | YesOrganization roles and shared-document view / edit permissions |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | Not documentedEach teammate runs local agents; isolated users on one shared execution server are not documented |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documentedLocal agents and provider usage; aggregate host compute budgets not documented |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | Not documentedWorktree creation; managed isolated compute provisioning not documented |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | Not documentedLocal terminal / extension system; managed Docker execution lifecycle not documented |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | YesSeveral sessions in one worktree or separate worktree sessions |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Manual setupDesktop auto-updates; native providers and local dependencies |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Manual setupReview files read / written and archive sessions; artifact deletion is manual |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Not documentedNot documented in reviewed sources |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | Not documentedDesktop app plus task CLI; standalone execution-controller binary not documented |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedAn integrated per-container journal covering these tool classes and verified installation outcomes is not established by the cited sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedDeletion triggered by all coding sessions exiting is not established by reviewed sources; timeouts and inactivity policies are different |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | YesSaved session records / transcripts and database backups; resume agent work |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | Durable transcripts / window state; process survival after desktop exit not documented |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | Not documentedFile/tool history is available; administrative permission and lifecycle audit events are not documented |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedNot documented in reviewed sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedNot documented in reviewed sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | Not documentedToken / tool counts by project; host CPU / RAM / disk accounting not documented |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedNot documented in reviewed sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | Project token activity; dollar charges and compute chargeback not documented |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Project trust plus native review, action / path / URL approvals |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | Not documentedNative permission modes; additional Nimbalyst OS sandbox not documented |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | Manual setupConfigure the host firewall yourself; native URL approval patterns are not an OS-wide egress boundary |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedNot documented in reviewed sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | Not documentedNative tool approvals and sensitive-path blocks are documented; OS-enforced shell filesystem containment is not established |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | CLAUDE.md / AGENTS.md guide behavior; project trust and action permissions govern tools |
Orca comparison
Agent IDE with local worktrees, arbitrary SSH hosts, a self-hosted server, CLI orchestration and experimental per-workspace cloud recipes. A worktree alone is not a container or VM boundary.
Sources: Orca: local, SSH, self-hosted server and experimental BYO cloud recipes · Orca: arbitrary SSH hosts, persistent remote terminals and forwarding · Orca: self-hosted servers, revocable pairing and reconnect · Orca: CLI harness support, worktrees, desktop/mobile and MIT source| Feature | devctrlSelf-hosted control plane | OrcaAgent IDE + SSH/self-hosted runtime |
|---|---|---|
| Software pricing | Free self-hosted software | Free/open-source application; model subscriptions and BYO cloud compute remain yours |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesRemote terminals survive client close/disconnect; managed server reattaches. Server restart and native conversation resume are distinct |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesArbitrary configured SSH hosts and self-hosted Orca servers |
| Run the control plane on your server | YesRun controller and browser gateway on your server | YesRun desktop as server or headless Orca serve on your machine; private-network pairing |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | YesDesktop and paired browser/mobile clients use a self-hosted runtime |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | LimitedSSH file downloads and recursive folder transfer when SFTP supports it; some actions are desktop-only |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Local/SSH use Git worktrees. Experimental recipes may create containers/VMs; worktrees are not infrastructure isolation |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedNot documented in reviewed sources |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Not documented in reviewed sources |
All features
| Feature | devctrlSelf-hosted control plane | OrcaAgent IDE + SSH/self-hosted runtime |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Free/open-source application; model subscriptions and BYO cloud compute remain yours |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Not documented in reviewed sources |
| Execution compute costs | Your local/remote compute bill | Not documented in reviewed sources |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | Desktop macOS/Windows/Linux plus mobile companion |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | MIT repository; CLI execution is broader than verified lifecycle support for every harness |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Not documented in reviewed sources |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Local, SSH and self-hosted server modes; experimental BYO per-workspace cloud recipes |
| Run agents on this computer | YesStart sessions on this computer; Docker required | YesAgents run on the desktop in local mode |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | Manual setupBring and pay your own provider; Orca does not sell managed VPS hosting |
| Run the control plane on your server | YesRun controller and browser gateway on your server | YesRun desktop as server or headless Orca serve on your machine; private-network pairing |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Local/SSH use Git worktrees. Experimental recipes may create containers/VMs; worktrees are not infrastructure isolation |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | YesSeparate Git worktrees for parallel tasks |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesRemote terminals survive client close/disconnect; managed server reattaches. Server restart and native conversation resume are distinct |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | LimitedOne desktop can drive multiple SSH targets; paired servers also exist. A combined devctrl-style fleet accounting view is not established |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesArbitrary configured SSH hosts and self-hosted Orca servers |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | Not documentedNot documented in reviewed sources |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | YesDesktop and paired browser/mobile clients use a self-hosted runtime |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedNot documented in reviewed sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedNot documented in reviewed sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedNot documented in reviewed sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedNot documented in reviewed sources |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedNot documented in reviewed sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedNot documented in reviewed sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedNot documented in reviewed sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedNot documented in reviewed sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedA separate frozen-session view with retained terminal snapshots and this S3 recovery workflow is not established by the reviewed sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedAn equivalent tree of agent-reported tasks and hook-recorded native steps for each session is not established by reviewed sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedAn owner-scoped inbox in which separate agent sessions exchange messages through a managed MCP tool is not established by reviewed sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | LimitedSSH file downloads and recursive folder transfer when SFTP supports it; some actions are desktop-only |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documentedNot documented in reviewed sources |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Not documentedNot documented in reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedNot documented in reviewed sources |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | Not documentedNot documented in reviewed sources |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | Not documentedNot documented in reviewed sources |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | Not documentedNot documented in reviewed sources |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | Not documentedNot documented in reviewed sources |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | Not documentedNot documented in reviewed sources |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedNot documented in reviewed sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedNot documented in reviewed sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | Not documentedNot documented in reviewed sources |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | LimitedSkills registry and update workflow documented; subject-based matching and weekly recommendation semantics are unverified |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | YesPer-provider usage and reset observations; unavailable values remain explicit |
| Continue on another subscription | YesEligible account + saved context | Not documentedNot documented in reviewed sources |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | Not documentedNot documented in reviewed sources |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | Not documentedNot documented in reviewed sources |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedNot documented in reviewed sources |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedNot documented in reviewed sources |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedNot documented in reviewed sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedThis comparison has not established equivalent per-account native model catalogs and default selection from the cited sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedShared per-user account display ordering across clients is not established by the cited sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | YesCLI orchestration, skills registry and MCP integration |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | Not documentedNot documented in reviewed sources |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedNot documented in reviewed sources |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | Not documentedNot documented in reviewed sources |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | Not documentedNot documented in reviewed sources |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | Not documentedNot documented in reviewed sources |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedNot documented in reviewed sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | LimitedTracked CLI orchestration and parent/child work documented; policy-equivalent cross-runtime validation unverified |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | Not documentedNot documented in reviewed sources |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | Not documentedNot documented in reviewed sources |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedNot documented in reviewed sources |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | Not documentedNot documented in reviewed sources |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | Not documentedNot documented in reviewed sources |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documentedNot documented in reviewed sources |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | Manual setupExperimental cloud recipes execute your provider create/suspend/resume/destroy lifecycle scripts |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | Not documentedNot documented in reviewed sources |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | Not documentedNot documented in reviewed sources |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Not documentedNot documented in reviewed sources |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Not documentedNot documented in reviewed sources |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Not documentedNot documented in reviewed sources |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | Not documentedNot documented in reviewed sources |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedAn integrated per-container journal covering these tool classes and verified installation outcomes is not established by the cited sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedDeletion triggered by all coding sessions exiting is not established by reviewed sources; timeouts and inactivity policies are different |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | Not documentedNot documented in reviewed sources |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | Remote runtime owns live processes; reconnect preserves workspace/tab state. This does not promise live process migration |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | Not documentedNot documented in reviewed sources |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedNot documented in reviewed sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedNot documented in reviewed sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | Not documentedNot documented in reviewed sources |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedNot documented in reviewed sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | Not documented in reviewed sources |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedCloud recipe diagnostic validation exists; equivalent installation/repository registry, AI checks and exact completeness contract are not established |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Not documented in reviewed sources |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | Not documentedNot documented in reviewed sources |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | Not documentedNot documented in reviewed sources |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedNot documented in reviewed sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | Not documentedNot documented in reviewed sources |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | Not documented in reviewed sources |
Paseo comparison
An open-source agent development environment with a local or self-hosted daemon, desktop/mobile/web clients, worktrees, schedules and agent orchestration.
Sources: Paseo: Getting started · Paseo: Local daemon and clients · Paseo: Schedules and heartbeats · Paseo: Agent orchestration| Feature | devctrlSelf-hosted control plane | Paseo |
|---|---|---|
| Software pricing | Free self-hosted software | Free, open-source Apache-2.0 software |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesAgents remain owned by the running daemon when a client disconnects |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesConnect to an existing remote daemon over SSH, relay or a direct connection; install it yourself |
| Run the control plane on your server | YesRun controller and browser gateway on your server | YesRun the daemon and bundled web UI on your server |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | YesThe daemon can serve the browser client for agent control |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Host processes and Git worktrees; optional daemon-wide Docker deployment |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Paired encrypted relay or direct daemon access with optional password; scope Docker mounts yourself |
All features
| Feature | devctrlSelf-hosted control plane | Paseo |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Free, open-source Apache-2.0 software |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Use your provider subscriptions or API billing; normal provider limits apply |
| Execution compute costs | Your local/remote compute bill | You supply the execution machine or server |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | macOS / Windows / Linux desktop; iOS / Android; web; CLI |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | Daemon with provider CLI adapters and a TypeScript SDK |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Agent development environment: workspaces, editor, terminals, diffs and browser |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Local or self-hosted daemon; optional encrypted relay and separate Hub |
| Run agents on this computer | YesStart sessions on this computer; Docker required | YesRun installed provider CLIs on the daemon machine |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | Not documentedHosted Hub routes events to your daemons; it is not documented as vendor-managed agent compute |
| Run the control plane on your server | YesRun controller and browser gateway on your server | YesRun the daemon and bundled web UI on your server |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Host processes and Git worktrees; optional daemon-wide Docker deployment |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | YesUse an existing directory or create a separate Git worktree and branch |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesAgents remain owned by the running daemon when a client disconnects |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesConnect to an existing remote daemon over SSH, relay or a direct connection; install it yourself |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | YesThe daemon can serve the browser client for agent control |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | YesDelegated workers appear in the Subagents track and report completion |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | Manual setupPersist mounted workspace and daemon directories in your Docker deployment |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | LimitedWorktree setup hooks and workspace scripts prepare task directories; agent CLIs are separate prerequisites |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | YesAccount quota windows, balances and reset times; unavailable reports stay explicit |
| Continue on another subscription | YesEligible account + saved context | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | LimitedAgents choose saved provider/model profiles using their delegation notes; no native before-spawn advisor is established |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | YesCLI, SDK and opt-in MCP tools manage agents, workspaces and workers |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | LimitedHub starts agents from GitHub, Slack and Discord events; this does not establish live messenger approval control |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Manual setupConfigure the bundled web UI, listening address, password and your reverse proxy |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | YesDelegate to configured providers/models and coordinate workers across reachable daemons |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | Manual setupDeploy the supplied daemon image and add agent CLIs in a child image; per-workspace container management is not established |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | LimitedReuse a local workspace or create a worktree; this is file isolation, not a managed container choice |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Manual setupInstall and authenticate provider CLIs on the execution host or in your child image |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | LimitedArchiving removes a managed worktree after its last workspace; a local project directory remains |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Manual setupInstall and maintain additional tools on the execution host yourself |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | LimitedNative history support depends on the provider; Antigravity retains the conversation but cannot replay its transcript |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | The daemon owns agents independently of connected clients; worktrees isolate files, not processes |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | Account quota reports do not measure per-agent token consumption or cost |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Paired encrypted relay or direct daemon access with optional password; scope Docker mounts yourself |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | LimitedDocker deployment isolates the daemon and agents together; it does not document a separate container for every session |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | Manual setupConfigure Docker mounts and host permissions yourself; agents can access their mounted code and credentials |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | Agent instructions and provider permissions do not establish separate per-user containers |
Superset comparison
A worktree-first IDE with a headless host daemon, remote relay, quota meters and cross-agent handoffs.
Sources: Superset model · Superset hosts · Superset remote · Superset accounts · Superset handoffs| Feature | devctrlSelf-hosted control plane | SupersetAgent workspace IDE |
|---|---|---|
| Software pricing | Free self-hosted software | Free local; Pro $20/user/month ($15 with annual billing); Enterprise custom |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesSessions keep running after client disconnect; execution host must stay online |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesInstall its headless host service on your server and connect through the relay; Pro required |
| Run the control plane on your server | YesRun controller and browser gateway on your server | LimitedSelf-host the terminal/host service; remote desktop/mobile access uses Superset’s relay |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedDesktop / iOS; browser console not documented |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedFile explorer; direct transfers not documented |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Local host processes/git worktrees; external Docker via lifecycle scripts |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesOrganization accounts and host owner/member access |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Manual setupProject setup/teardown can run Docker; devcontainer management not documented |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Host user permissions/worktree boundaries; relay/host access authorization |
All features
| Feature | devctrlSelf-hosted control plane | SupersetAgent workspace IDE |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Free local; Pro $20/user/month ($15 with annual billing); Enterprise custom |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Your provider subscriptions/API bill |
| Execution compute costs | Your local/remote compute bill | Your local/remote compute bill; optional Cloud terms apply separately |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | macOS, experimental Linux desktop; iOS/iPadOS Pro; headless hosts |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | TypeScript/React/Electron desktop; host/PTY services |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Task worktree IDE plus remote host/agent management |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Local-first desktop, dedicated headless hosts and vendor relay |
| Run agents on this computer | YesStart sessions on this computer; Docker required | YesStart sessions on this computer |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | YesCloud workspaces in limited early access |
| Run the control plane on your server | YesRun controller and browser gateway on your server | LimitedSelf-host the terminal/host service; remote desktop/mobile access uses Superset’s relay |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Local host processes/git worktrees; external Docker via lifecycle scripts |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | YesManaged Git worktrees locally; separate clones in Cloud |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesSessions keep running after client disconnect; execution host must stay online |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | Not documentedDevice filter switches between host workspaces; one combined multi-host session list is not documented |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesInstall its headless host service on your server and connect through the relay; Pro required |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | YesShows working / waiting-for-input states for supported agent sessions |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | Not documentedDesktop / iOS; browser console not documented |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedNot documented in reviewed sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedNot documented in reviewed sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedNot documented in reviewed sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedNot documented in reviewed sources |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedNot documented in reviewed sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedNot documented in reviewed sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedNot documented in reviewed sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedNot documented in reviewed sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedNative session forks and handoffs are documented; automatic compaction after completed work is not established |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedA separate frozen-session view with retained terminal snapshots and this S3 recovery workflow is not established by the reviewed sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedAn equivalent tree of agent-reported tasks and hook-recorded native steps for each session is not established by reviewed sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedAn owner-scoped inbox in which separate agent sessions exchange messages through a managed MCP tool is not established by reviewed sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedFile explorer; direct transfers not documented |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documented |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Manual setupConfigure host paths/mounts in project setup scripts; a managed remote-folder grant is not documented |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Manual setupProject setup/teardown can run Docker; devcontainer management not documented |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | YesHost worktrees, local state and terminal history; restore committed archived work |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | LimitedReuse project setup/run/teardown scripts; prepared compute-image reuse is not established |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | Manual setupHost credentials and project setup; remote host sharing can expose them |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | LimitedForward ports started by the selected workspace; a centrally granted shared-service catalog is not documented |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | Not documentedResource monitoring documented; per-user hard enforcement not documented |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedNot documented in reviewed sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedNot documented in reviewed sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | YesProject setup scripts and onboarding tool/account checks |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | YesClaude / Codex meters per host |
| Continue on another subscription | YesEligible account + saved context | LimitedRelaunch Claude with another profile and saved history; Codex account continuation is not established |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | YesContinue the task with another agent through supported native fork/context handoff |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | Not documentedModel overrides; task advice not documented |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documented |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedCost analytics; routing budgets not documented |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedNot documented in reviewed sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedThis comparison has not established equivalent per-account native model catalogs and default selection from the cited sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedShared per-user account display ordering across clients is not established by the cited sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | YesAgents create/manage workspaces and send prompts through MCP/CLI/SDK tools |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | Not documentedHost access; project resource grants not documented |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedAgent approvals; project-grant flow not documented |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | Not documentedManagement MCP is documented; a managed external-tool allowlist is not |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | Not documented |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | LimitedSlack starts tasks/agents; interactive session prompts and approval replies are not established |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedNot documented in reviewed sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documented |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | LimitedTeammate host access exposes its reachable terminals; session-specific view/control grants are not documented |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | LimitedOrganization repositories have task worktrees; host access does not isolate each member’s files/credentials |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesOrganization accounts and host owner/member access |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | YesHost owner/member permissions; Enterprise SSO/SCIM |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | Not documentedHost sharing exposes reachable files/terminals; private per-user containers not documented |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documentedResource views; aggregate budgets not documented |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | Manual setupWorktrees / Docker setup scripts |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | Manual setupProject Docker setup scripts |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | Manual setupShared host / external containers |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Manual setupAgent CLIs on host PATH |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Manual setupWorkspace teardown scripts |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Manual setupInstall and update tools yourself; no managed context-tools workflow documented |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | Not documentedDesktop app + headless host |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedAn integrated per-container journal covering these tool classes and verified installation outcomes is not established by the cited sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedDeletion triggered by all coding sessions exiting is not established by reviewed sources; timeouts and inactivity policies are different |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | YesUnexpected agent termination auto-resumes supported saved session IDs |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | PTY daemon survives app/host service restart; rebooted agents resume history |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | YesEnterprise audit log; plan requirement applies |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedNot documented in reviewed sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedNot documented in reviewed sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | LimitedLocal process CPU/RAM view; remote host consumption is not aggregated |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedNot documented in reviewed sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | API-equivalent log costs by model/workspace; not subscription bill |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Host user permissions/worktree boundaries; relay/host access authorization |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | Manual setupUse a dedicated host/container; no universal app-level OS sandbox claimed |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | Not documentedEnterprise IP restrictions; workload egress enforcement not documented |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedNot documented in reviewed sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | Manual setupHost permissions; file editor can follow paths outside a worktree |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | Agent hooks/instructions do not turn worktree isolation into an OS boundary |
T3 Code comparison
An open-source control surface for the agent CLIs on your machines, with desktop, web and mobile clients. Its server runs on your computer, a server or a VPS; agents work in project directories or Git worktrees on that host.
Sources: T3 Code: Agent harness control surface · T3 Code: Pricing · T3 Code: Remote access · T3 Code: Project settings| Feature | devctrlSelf-hosted control plane | T3 CodeAgent control surface + your own servers |
|---|---|---|
| Software pricing | Free self-hosted software | Free, open-source MIT software; T3 Connect links up to 3 machines, more with a paid T3 Chat plan |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesSessions keep running after client disconnect; execution host must stay online |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesDesktop-managed SSH starts or reuses a server on Linux or Apple Silicon hosts; paired servers also connect |
| Run the control plane on your server | YesRun controller and browser gateway on your server | YesRun the T3 Code server as a background service on your server |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | YesThe local or hosted web app sends prompts to the same server threads |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | LimitedAttach files to messages; browser downloads are offered to your device. Workspace archive downloads are not established |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Host processes in project directories or Git worktrees; container isolation is not documented |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedPairing links and sessions carry permission scopes; application user accounts are not established |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Pairing links and scoped session tokens; T3 Connect sets up a Cloudflare tunnel and does not proxy sessions |
All features
| Feature | devctrlSelf-hosted control plane | T3 CodeAgent control surface + your own servers |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Free, open-source MIT software; T3 Connect links up to 3 machines, more with a paid T3 Chat plan |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Use your provider subscriptions or API keys; T3 Code does not resell tokens |
| Execution compute costs | Your local/remote compute bill | You supply the computer, server or VPS that runs the T3 Code server |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | macOS / Windows / Linux desktop and CLI; iOS / Android; web app |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | Server that controls installed provider CLIs; Electron desktop, web and mobile clients |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Agent control surface: threads, diffs, terminals, browser tabs and pull requests across providers |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Local or self-hosted server; reach it through T3 Connect, LAN, Tailscale or SSH |
| Run agents on this computer | YesStart sessions on this computer; Docker required | YesAgents run wherever the T3 Code server runs, including this computer |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | Not documentedAgents run on your machines; vendor-managed execution compute is not established |
| Run the control plane on your server | YesRun controller and browser gateway on your server | YesRun the T3 Code server as a background service on your server |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Host processes in project directories or Git worktrees; container isolation is not documented |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | YesEach agent thread writes to its own branch; managed Git worktrees have cleanup policies |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | YesSessions keep running after client disconnect; execution host must stay online |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | LimitedClients connect several environments and search threads across them; a combined fleet accounting view is not established |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesDesktop-managed SSH starts or reuses a server on Linux or Apple Silicon hosts; paired servers also connect |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | YesWorking, approval and question states; push notifications when an agent finishes, fails or needs input |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | YesThe local or hosted web app sends prompts to the same server threads |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedSubagent approvals and working state appear in the parent thread; an active subagent count is not established |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedClaude auto-compaction follows a token threshold; compaction after completed work is not established |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | YesAgents start, message and wait on threads, including on linked environments (beta) |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | LimitedAttach files to messages; browser downloads are offered to your device. Workspace archive downloads are not established |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | LimitedCredential fields can reference 1Password and agents request secrets privately; per-scope runtime delivery is not established |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | LimitedProject setup scripts and worktree submodule initialization; provider CLIs are host prerequisites |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | YesPooled subscription limit windows, remaining quota and reset times per provider in the Usage view |
| Continue on another subscription | YesEligible account + saved context | LimitedSwitch a Codex thread to another compatible account manually; T3 Code deliberately has no automatic account switcher |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | YesSwitch providers in a thread with a budgeted context handoff; reasoning and tool-call state are not copied |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedUsage estimates API-equivalent cost by model; per-project routing budgets are not established |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | YesMCP server lets agents read and manage projects and threads with a chosen access mode |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | LimitedA machine-wide managed policy sets defaults, locks and permission-mode caps |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | Not documentedWebhook tasks start runs from other services; messenger bot control is not established |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | LimitedChoose T3 Connect, LAN, Tailscale HTTPS or another route; the hosted web app needs an HTTPS endpoint |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | Yesdelegate_task hands work to a linked environment; the parent thread wakes with the result |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | LimitedGuest links (beta) let one person follow a thread and propose messages you send or dismiss |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | Not documentedPairing links and sessions carry permission scopes; application user accounts are not established |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | LimitedPairing and session tokens take permission scopes; outside agents get read-only to full-access modes |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Manual setupInstall and authenticate provider CLIs on the server host; T3 Code updates itself through its own command |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | LimitedWorktree cleanup policies remove inactive, settled, merged or deleted-thread worktrees and report kept ones |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | YesInstall the t3 command with its install script, or run it once with npx |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | LimitedThreads and queued messages survive server restarts; continuing may start a fresh provider session with a context handoff |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | The server owns agents and threads; phones, browsers and desktops are remotes to it |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | Estimated API-equivalent cost by model and environment from provider history; not your subscription bill |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Pairing links and scoped session tokens; T3 Connect sets up a Cloudflare tunnel and does not proxy sessions |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | Not documentedContainer or VM isolation for agents is not established; provider permission modes are separate |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | Not documentedPermission modes request approval through the provider; an enforced filesystem boundary outside it is not established |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | Permission modes rely on each provider’s enforcement; a managed policy caps modes but does not create containers |
Warp comparison
Terminal with its own Warp Agent, native CLI-agent support and browser session sharing. Its Automation Platform adds hosted and Enterprise self-hosted execution.
Sources: Warp overview · Warp CLI agents · Warp architecture · Warp self-hosting| Feature | devctrlSelf-hosted control plane | WarpAgentic terminal + cloud platform |
|---|---|---|
| Software pricing | Free self-hosted software | Free terminal + metered AI; Build from $20/month; Enterprise custom |
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | LimitedCloud runs continue after disconnect; native CLI Remote Control requires the original host/session to remain running |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesSSH terminal; Enterprise workers or Oz CLI on your existing infrastructure |
| Run the control plane on your server | YesRun controller and browser gateway on your server | NoWarp operates the control plane; Enterprise self-hosts execution workers |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | YesPublished agent sessions and cloud dashboard; web / mobile viewing and steering |
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedCloud snapshots / artifacts; general direct file upload / archive download workflow not documented |
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Local/SSH host processes; Cloud containers/VMs; Enterprise direct workers share the host OS |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesWarp users / teams, personal or agent API keys and integration identities |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedDocker environment / runner images; Dev Container specification not documented |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Agent permission profiles + cloud isolation; Enterprise enforced policies / own-host controls |
All features
| Feature | devctrlSelf-hosted control plane | WarpAgentic terminal + cloud platform |
|---|---|---|
| Product and cost | ||
| Software pricing | Free self-hosted software | Free terminal + metered AI; Build from $20/month; Enterprise custom |
| Model usage costs | Your provider subscriptions/API bill; optional Decisions advisor calls cost extra | Native CLIs: Your provider subscriptions/API bill. Warp Agent: credits or configured inference |
| Execution compute costs | Your local/remote compute bill | Cloud compute / platform credits; self-hosted hardware costs are yours |
| Client platforms | Terminal + web; macOS + Windows desktop; Linux build target; Android ARM64 preview. Clients are source-built; Android verified in an emulator. No iOS app. | macOS / Windows / Linux; desktop/mobile web browsers |
| Implementation | TypeScript/Node controller and agents; one shared React App for TUI, web, desktop and Android; SQLite; Tauri desktop/mobile shells. Shared built-in interface and native onboarding sources in 14 languages; native binaries need a new build | Rust client / native UI; MIT UI crates, AGPLv3 remainder |
| Primary product model | Self-hosted control plane for persistent agent environments and team infrastructure | Agentic terminal with Warp Agent and native CLI support; cloud automation platform |
| Infrastructure ownership | ||
| Local, self-hosted and vendor Cloud model | Local or self-hosted controller; SSH or HTTPS/WSS clients; optional local Docker, cloud hosts and capability-gated sandbox adapters | Local / SSH terminal; hosted cloud agents; Enterprise self-hosted workers |
| Run agents on this computer | YesStart sessions on this computer; Docker required | YesStart sessions on this computer |
| Use vendor-managed execution compute | YesOptional cloud-VM and sandbox packages; API-only adapters cannot launch Docker-managed sessions; provider billing and limits apply | YesWarp-hosted cloud sandboxes; plan-dependent resources / concurrency |
| Run the control plane on your server | YesRun controller and browser gateway on your server | NoWarp operates the control plane; Enterprise self-hosts execution workers |
| Execution isolation | ||
| Execution isolation boundary | Managed Docker containers; optional sandbox provider boundary | Local/SSH host processes; Cloud containers/VMs; Enterprise direct workers share the host OS |
| Give parallel tasks separate checkouts | YesSeparate working files, Git index and branch for every new coding session, including in a shared container; native worktrees or a shared-object clone on older Git. Existing sessions retain their paths | YesCloud per-run containers / VMs; self-hosted direct mode shares host OS |
| Sessions | ||
| Keep running after client disconnect | YesSessions keep running after client disconnect; execution host must stay online | LimitedCloud runs continue after disconnect; native CLI Remote Control requires the original host/session to remain running |
| One session list across execution hosts | YesProjects and Mosaic list sessions from enrolled hosts together | Not documentedCloud runs and SSH terminals are available; one combined native-agent session list is not documented |
| Use your own remote execution servers | YesConnect your existing execution hosts over SSH | YesSSH terminal; Enterprise workers or Oz CLI on your existing infrastructure |
| Identify sessions waiting for input | YesShows working / waiting-for-input states for supported agent sessions; Choose questions, permission requests, idle or exhausted quota in Settings; Projects counts visible matches and group tabs show a hand. Hidden sessions are excluded; sound is separate. optional local sounds for explicit questions, permissions and new MCP approvals while the console is open; off by default | YesAgent progress / tool activity, cloud run status and native CLI notifications |
| Send session prompts from a browser | YesOpen the web console and type in the same live session | YesPublished agent sessions and cloud dashboard; web / mobile viewing and steering |
| See actual stages while opening a session | YesTerminal and browser show reported account, access, workspace, import and attachment stages with elapsed time; cancel without stopping the persistent agent | Not documentedNot documented in reviewed sources |
| Name sessions without changing their native history | YesRename the visible session or review a local title suggestion without changing conversation IDs or prompting the agent | Not documentedNot documented in reviewed sources |
| Capture periodic progress summaries in session headers | YesServer-side Claude /btw status questions every 15 minutes by default, with version/input/viewer guards, capture date and previous answer retained on failure. Uses agent quota. Codex automation is disabled because /side changes the visible terminal; other agents require verified safe capture | Not documentedNot documented in reviewed sources |
| Show active native subagent counts | YesClaude/Codex native telemetry reports active children, models, own child goals and goal-scoped checklist progress where supported; unavailable or stale observations are shown as unknown | Not documentedNot documented in reviewed sources |
| Open a web server running in a session from the browser | YesOpt-in per session: the agent binds its HTTP server to a stable preview port and the owner opens a one-use link on a per-session origin. Only that container port is reachable; WebSocket upgrades work. The listener is local by default; reach it over an SSH port forward or a per-port origin | Not documentedNot documented in reviewed sources |
| Mark what each session is doing | YesActivity tags from structured state (waiting input, needs approval, blocked, quota, frozen, outdated CLI, working) plus agent-declared work kinds and milestones and Git facts such as unpushed commits; nothing is inferred from free text | Not documentedNot documented in reviewed sources |
| Separate main-session and subagent work | YesTask Graph attributes every task to the main session or the subagent that did it, with per-agent codes, lanes and scoping | Not documentedNot documented in reviewed sources |
| Show conversations between agent sessions | YesTask Graph lists messages, questions, replies, delegations and handoffs between sessions with their status; bodies are reduced to a redacted subject line | Not documentedNot documented in reviewed sources |
| Auto compact after completed work | YesManaged Claude/Codex/Antigravity: submit /compact server-side after an explicit completed-task report and native turn end, even with the client offline. Off by default; Global → Project → Session policy, idle delay and cooldown. Questions/permissions, active or unknown children, attached terminals and quota waits block it. Native context-limit compaction is separate. | Not documentedWarp Agent has /compact and /compact-and; completed-task automation for native CLI sessions is not established |
| Freeze an agent and retain its terminal snapshot | YesStop one agent and retain its screen, native conversation and checkout in Frozen. Thaw prepares the container before resuming the same account/profile; idle containers stop. Temporary compute requires verified encrypted S3 storage. Permanently remove a confirmed stopped card with its exact Freeze ID; retained files/history/archives and a retry receipt remain. | Not documentedA separate frozen-session view with retained terminal snapshots and this S3 recovery workflow is not established by the reviewed sources |
| Follow each agent’s tasks and steps as a tree | YesTask Graph combines tasks agents report through the devctrl MCP tool with native Claude hook steps: classified shell commands, checklists and named subagents, with sanitized agent-written descriptions, start times and running durations. Raw commands are never recorded. Projects Activity shows current work in the mosaic grid. Reports are unverified progress claims; Codex plans and subagents are not captured. | Not documentedAn equivalent tree of agent-reported tasks and hook-recorded native steps for each session is not established by reviewed sources |
| Let agent sessions message each other | YesClaude, Codex, Antigravity and Pi sessions of the same owner in one Project exchange messages through devctrl MCP tools, across enrolled hosts. Messages arrive in an inbox inside an untrusted-peer envelope and are never typed into a terminal; idle Claude/Codex sessions can receive a body-free wake line. Not yet exercised end to end with real agents; team members and other servers are not included. | Not documentedAn owner-scoped inbox in which separate agent sessions exchange messages through a managed MCP tool is not established by reviewed sources |
| Files and storage | ||
| Upload inputs and download artifacts | YesUpload and download through Exchange in web/TUI; no scp commands | Not documentedCloud snapshots / artifacts; general direct file upload / archive download workflow not documented |
| Automatically sync a folder both ways | YesTwo-way Unison through the optional SSH client. Per-user HTTPS/WSS synchronization is a preview with an unresolved clean-restart limitation. Each selected folder is a whole-folder read/write grant | Not documentedWorkspace handoff snapshots; general bidirectional folder sync not documented |
| Attach folders from local/remote sources | YesHost-local / SSH / S3 bucket folders with MCP read/write grants. S3 uses the gateway. Shell mounts: host-local/pre-mounted sources in a new dedicated container | Manual setupAgent permission profile and configured self-hosted container mounts |
| Save workspaces and session history to S3 | YesOvernight saves encrypted, verified workspace and conversation archives to S3-compatible storage, with an index of saved work. Restore on the same enrolled host when you return, or run pnpm overnight:restore to place saved repository and native-history trees in new isolated directories on another host without resuming sessions. Runtime credentials are excluded. Import local Claude Code and Codex sessions directly into devctrl Cloud or your own S3 bucket. Connect multiple buckets and choose the default for new imports. See the project and repository for each import and download the original history. Continue an imported session on a managed Boat sandbox using your subscription. | Not documentedA supplied S3 export/restore workflow covering workspaces and native coding-session history is not established by reviewed sources |
| Environment management | ||
| Give one agent multiple repositories | YesSelect existing repositories or import an entire accessible GitHub organization; save launch groups with independent checkouts. Operator Control Room suggestions stay within one verified organization. Matching runtime, Git grants and Internet policies required | Not documentedNot documented in reviewed sources |
| Create environments from devcontainer.json | YesRepo devcontainer config for operator; managed template for tenants | Not documentedDocker environment / runner images; Dev Container specification not documented |
| Keep workspace files when compute stops | YesDisk-backed workspaces/history; disposable providers have explicit lifetimes | LimitedCloud transcripts/artifacts and optional snapshots persist; the live sandbox filesystem is torn down |
| Reuse prepared execution environments | YesManaged runtime templates and cached prepared image variants | YesReusable Docker environments / setup and runner definitions |
| Deliver secrets to selected runtime scope | YesEncrypted controller vault; scoped runtime file/broker delivery | YesPersonal / team secrets injected at runtime; per-run secret allowlists |
| Publish shared services to selected users/projects | YesAdmin publishes resources; users grant them to owned projects | YesTeam MCP, secrets and integrations; self-hosted workers reach private services |
| Enforce CPU/RAM/storage resource limits | YesAggregate CPU/RAM/PID cgroups and XFS/ext4 quotas require verified host delegation and quota prerequisites. Finite running-container CPU/RAM requests preserve user budgets; approval stays separate from verified Docker/kernel/parent limits and session preservation | YesPlan-bounded vCPU / RAM shapes and cloud concurrency; host backend limits configurable |
| Choose a workspace storage base per user/host | YesAdministrators choose a provisioned storage base per user/host for new workspaces; retained workspace paths do not move | Not documentedNot documented in reviewed sources |
| Pin a user's containers to selected CPUs | YesVerified cgroup-v2 CPU sets apply across a user's containers; updates preserve sessions. Affinity is not exclusive-core reservation | Not documentedNot documented in reviewed sources |
| Prepare tools and dependencies automatically | YesPrepare checkouts, selected agent/Git tools and images. Custom GitLab hosts/ports configure controller API access. Restricted workspace Git needs an approved public host on ports 80/443; private/custom-port routes are not provisioned automatically | YesClone repos, apply Docker environment, run setup and inject secrets |
| Preserve repository technology evidence by subject | LimitedBounded static ecosystem parsers, component/evidence snapshots and match/non-match/unknown rules; dynamic syntax remains partial | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Refresh technology-aware skill recommendations | LimitedCurated language/technology JSON links match bounded manifest evidence. Project defaults and repository exclusions preserve manual choices; pinned bundle installation is explicit. Daily cached catalog metadata stays separate from installed files, compatibility and execution evidence; refresh never activates skills | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Accounts and models | ||
| Show agent subscription quota and resets | YesTrack remaining quota and reset times for Claude/Codex. Keep Antigravity limits separate by model group. Show Claude free-reset offers and Codex reset credits when officially reported. Confirm supported Codex resets in the console; follow account-specific instructions for Claude. | Not documentedWarp credit usage; native Claude / Codex quota meters not documented |
| Continue on another subscription | YesEligible account + saved context | Not documentedOwn inference / supported subscriptions; quota-triggered native account continuation not documented |
| Continue across Claude / Codex / Antigravity | YesSupported Claude/Codex imports; bounded task handoffs including Antigravity. Original session retained | Not documentedLocal / cloud runtime handoff; cross-native Claude / Codex transcript transfer not documented |
| Recommend a model for a delegated task | YesHosted or self-hosted Decisions advisor selects cheaper eligible subagent models; forecasts cost | LimitedWarp Agent can choose models automatically; native Claude/Codex delegated-task advice is not established |
| Apply model advice in native subagent hooks | YesClaude hooks and supported Codex launches; inherited-history Codex forks keep their model | Not documentedWarp orchestration; native CLI model-advice hook not documented |
| Track routing budgets by project | YesBudgets + estimated model costs | Not documentedCost-efficient Auto routing / team spend caps; per-project routing budgets not documented |
| Choose a proxy route per agent account | YesOptional HTTP/SOCKS/existing VPN-gateway adapters in restricted Docker workspaces (rootful 28+ or rootless 29.9.1+); changes apply at the next provider process start | Not documentedNot documented in reviewed sources |
| Choose native model defaults per account | YesChoose current installed Claude/Codex/Antigravity model catalogs or native default. Antigravity checks an opt-in reserve chain on the server. Verified native quota refusals skip the failed group even before cached usage updates. Idle selection preserves the conversation and read-only public viewers; unknown reserve quota blocks selection. A verified quota-interrupted task receives one continuation prompt after selection is confirmed; uncertain input is not replayed. Claude accounts can instead use a metered OpenRouter, Vercel AI Gateway or LiteLLM gateway with a model allowlist, recorded per-request cost and an optional monthly budget; these have no subscription quota windows. | Not documentedThis comparison has not established equivalent per-account native model catalogs and default selection from the cited sources |
| Keep your preferred account order across clients | YesManual account order saved per user for detailed and compact Accounts in TUI, web and desktop. Account and nested session rows move together. | Not documentedShared per-user account display ordering across clients is not established by the cited sources |
| Access and integrations | ||
| Let agents manage product sessions/resources | YesAgents call devctrl MCP management tools and bundled skills. Optional Public Project MCP, off by default, lets external OAuth clients read granted Projects, sessions and task graphs or send text to a running session; Control Room through the remote console remains the primary remote path | YesWarp agent terminal tools, MCP and cloud CLI / API orchestration |
| Grant service access per project | YesConnect resources once; grant selected projects routed endpoints and scoped runtime credentials | YesScoped cloud secrets, environment / team integrations and MCP server permissions |
| Review project access requestsRelease video on YouTube: Approval notices in Main (devctrl 0.1.171) | YesReview exact actions and execute approved plans sequentially on the server. Add system, Project-group and repository approval policies; confirm editable general intents for legal repository or Project-group scopes, inspect and revoke them. Owner/account/team ceilings remain binding. Intent, approval, queued delivery and verified technical outcome are separate; uncertain writes are not replayed | Not documentedInteractive command / MCP approvals; central project-grant delivery not documented |
| Allow/deny specific MCP tools or methods | YesManaged MCP broker enforces allowed tools/methods for the granted scope | LimitedWarp profiles control MCP servers; per-method filtering for native CLI sessions is not established |
| Reuse access-policy templates | YesSave access/rate templates and apply them to groups, repositories or sessions | YesAgent profiles; Enterprise role-based admin-enforced settings |
| Control sessions through messenger bots | YesTelegram, Slack and Teams support prompts and approvals after bot setup | LimitedSlack can trigger Cloud agents; native CLI session prompting and approval replies are not established |
| Save a public web access provider | YesSave Automatic, Cloudflare, localhost.run, ngrok or your HTTPS origin for the next web start; ngrok credentials/agent and own-domain DNS/TLS are external setup | Not documentedNot documented in reviewed sources |
| Cap MCP calls per repository/session | YesManaged broker enforces configured scoped rate limits and reusable templates | Not documented |
| Delegate tracked work across agent runtimes | LimitedPersisted isolated Claude/Codex tasks, shared admission and MCP results. Exact account must be logged in on the target host; managed fresh-VM tasks wait for explicit account preparation. Automatic native credential transport is incomplete | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Collaboration and tenancy | ||
| Share a live session (view / control) | YesSession owner grants view/control to server users or guests with a revocable browser link; one active input lease | YesPublished sessions: view-only or edit access for steering, in browser/desktop |
| Separate user environments for the same Git repository | YesUsers import the same Git URL into private Project records; separate user-scoped checkouts/containers | LimitedCloud runs clone team repositories into separate per-run environments; local/SSH user isolation is your setup |
| Application-level user identities | YesPrivate users; password sign-in and remembered devices in TUI, browser and app. Invited/active/disabled status, invitations of up to 7 days for new users only, reset access, disable/enable and an effective-access view | YesWarp users / teams, personal or agent API keys and integration identities |
| Role-based resource authorization | YesOperator/admin/user roles plus ownership checks at API and host; only the server owner manages other administrators, and every controller route has a checked access policy | YesEnterprise Owner / Admin / Member; session view / edit access |
| Isolate users on shared execution servers | YesDifferent users have separate containers, HOME and user-scoped repository/runtime paths; existing mounts stay in place | YesAccount-scoped cloud run access and separate per-run sandboxes |
| Preserve required team review skills in lower scopes | LimitedTeam/template/project/repository contracts preserve required and locked constraints; configured policies and compatible pinned bundles are required | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Teamwork | ||
| Cap total compute allocated to each user | YesAggregate launch budgets; hard enforcement needs verified host support | Not documentedTeam credit caps / per-run shapes; aggregate per-user CPU / RAM / disk budgets not documented |
| Compute and maintenance | ||
| Provision isolated compute for a workspace | YesCreate an isolated Docker workspace or a sandbox through a configured provider adapter | YesCloud environment provisioning; Enterprise worker backends |
| Manage Docker environments for agents | YesCreate, reuse, rebuild and remove managed agent containers | YesDocker-based cloud environments / Linux runners; Enterprise Docker workers |
| Reuse an environment or create a separate one | YesLaunch in a compatible same-owner container or create a separate one in the session wizard | YesLocal terminals / shared sessions; cloud runs use separate sandboxes |
| Upgrade CLI tools in selected workspaces | YesUpdate managed CLIs inside containers; running sessions retain launched versions. Stopped/Frozen cards show the next-start target, including matching rebuilt images. Rebuild images separately | Manual setupMaintain environment images/setup commands; a scoped native CLI workspace-upgrade action is not documented |
| Delete session-declared files/containers centrally | YesReview managed disk sizes and declared files/containers. Clean ready task artifacts, configure empty-container shutdown and optional hourly host retention policies | Not documentedCloud sandbox teardown; session-declared artifact cleanup manifests not documented |
| Manage optional context tools in containers | YesRTK OSS, sgrep, ast-grep, Repomix and Graphify: exact scoped selections, separate image/container application, managed-tool removal, per-tool progress, all-session guide/notice receipts and explicit failed-step retry; usage is unmeasured | Not documentedNot documented in reviewed sources |
| Install a standalone server binary | NoInstallation uses a Docker control-plane stack | Not documentedAgent/worker CLI binaries exist; they do not provide a standalone self-hosted control-plane binary |
| Overnight: shut down the server when you’re away | YesEnable Overnight before you leave. Keep agents working, then shut down the server once every session has finished or is waiting for you. Your client can stay offline. Available through hosting adapters with shutdown support. | Not documentedAn opt-in server shutdown mode triggered by every coding session finishing or waiting for its absent user is not established by reviewed sources |
| Inspect actual tooling changes per container | YesPaged container history for agent CLIs, context tools and observed runtime tools. Actual before/after versions, requested targets, source, actor and outcome stay separate; first observations and missing versions are explicit. | Not documentedAn integrated per-container journal covering these tool classes and verified installation outcomes is not established by the cited sources |
| Remove disposable compute after sessions finish | YesBoat workspaces can be deleted automatically when every session has exited. Choose this lifetime when creating the sandbox. Waiting or detached sessions do not trigger deletion; provider expiry still applies. | Not documentedDeletion triggered by all coding sessions exiting is not established by reviewed sources; timeouts and inactivity policies are different |
| Operations and accounting | ||
| Resume conversation history after host restart | YesRetained native session IDs support resume; running memory is not checkpointed | Not documentedCloud transcripts / snapshots can resume; automatic host-reboot process recovery not documented |
| Persistence mechanism and process boundaries | Controller disconnect leaves host provider processes alive; reboot uses resume | Cloud runs execute independently; local Remote Control sync ends when session closes |
| Record authorization and operation audit events | YesPaged operation events and network/MCP attempts by project, repository and session. Unique-target counts show allowed/denied and upstream failures separately. HTTP/CONNECT coverage with observed workspace/DinD origins; shared-container traffic stays shared | LimitedCloud run/tool history and costs; administrative authorization-change audit is not established here |
| Export visible operation events as redacted JSONL | YesExport retained operation events or network/MCP attempt metadata as scoped redacted JSONL; no credentials, URL queries or MCP arguments/results | Not documentedNot documented in reviewed sources |
| Stream operation events to a log collector | YesAuthenticated resumable WebSocket feeds, operator-only controller stdout and private server Unix sockets. Filter operation/network/MCP metadata by owner and project; saved cursors and bounded backpressure work with clients closed | Not documentedNot documented in reviewed sources |
| Show workspace CPU/RAM consumptionRelease video on YouTube: devctl and the rest of the Host on the Hosts screen (devctrl 0.6.0) | YesHost/container CPU, RAM and disk I/O history with Dashboard timelines and interval totals; allocation budgets are separate | Not documentedCompute credits and run sizing are reported; live workspace CPU/RAM consumption is not documented |
| Warn agents about measured RAM pressure | YesHost/controller RAM monitoring continues without the client; warnings wait for safe input. Current managed process metadata is required; no automatic process kills | Not documentedNot documented in reviewed sources |
| Workspace/project cost reporting | Project activity: native tokens without requiring Git, commits and line deltas by day/week/month, recorded 91-day totals, metric-specific Last and per-user hidden repositories; Claude account records add provider-week project/session/model attribution with cache and subagents; recorded token shares, Decisions forecasts and sandbox usage are separate from subscription billing | Run cost records and user / team credit billing; team spend caps |
| Enforce review against an immutable subject | LimitedServer gate binds subject and effective contract and blocks missing/unknown required evidence. Workflow settings expose user defaults, project overrides, triggers and per-account admission. Automatic hooks are limited to verified native local Git contexts; ordinary global/system and managed-host contexts require manual integration. Trusted identity and remote required-status configuration remain necessary. Review cycles, reviewer models, verdicts and top findings appear in the session's Task Graph | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Export versioned review outcomes and coverage | LimitedImmutable JSON attempt/history and bounded JSONL export; unknown execution/provenance stays explicit | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Run deterministic and account-backed diagnostic checks | LimitedCurrent problems show affected target, observed/expected state, impact and a repair, settings link or recheck. Safe fixes need trusted completion plus fresh verification; history and optional analysis stay separate. Live optional Claude analysis is unverified; Codex AI diagnostics unsupported | Not documentedEquivalent workflow is not established by the reviewed primary sources |
| Security boundaries | ||
| Security model | Container/tenant separation, supported Docker egress enforcement and managed routes; agent rules remain instructions | Agent permission profiles + cloud isolation; Enterprise enforced policies / own-host controls |
| Isolate execution with containers/VMs/OS sandbox | YesManaged Docker boundary or optional sandbox provider; same-owner shared containers share a trust boundary | LimitedCloud containers/VMs isolate execution; Enterprise direct-host workers do not add that boundary |
| Enforce workspace Internet-egress restrictionsRelease video on YouTube: Gateway health in devctl hosts (devctrl 0.3.0) | YesNew/rebuilt restricted workspaces use an isolated-network allowlist gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. System/user/project-group/repository packages inherit approved team templates; explicit repository policies require opt-in adoption. Public80/443 policy and gateway acknowledgement stay separate; required model/auth hosts remain allowed. Legacy runtimes need rebuild; older rootless daemons and Boat-native networking are unsupported | Manual setupSelf-hosted host / Kubernetes network policy; direct agents inherit host access |
| Screen text agents send each other before delivery | YesA bounded deterministic prescreen flags code, binaries and encoded payloads; Decisions or Control Room review suspicious text, and high-risk entry points such as delegated tasks and session input are reviewed even when clean. Holds wait for a person | Not documentedNot documented in reviewed sources |
| Enforce filesystem boundaries outside agent prompts | YesTenant mounts/ownership and folder gateways; native access outside grants is distinct | LimitedCloud container/VM filesystem boundary; direct-host workers need your host/mount restrictions |
| Instructions versus enforced access boundary | Approval policy text and confirmed task intents guide bounded review; broker/mount/ownership enforcement and actual technical fulfillment remain separate | Agent rules guide behavior; profiles / backend isolation enforce tool boundaries |
Yes: a supplied workflow. Limited: the stated edition or runtime. Manual setup: your commands, scripts or infrastructure configuration. Red marks unavailable or unverified support; each answer links to the evidence. Pricing, platforms and architecture are shown as facts.
THE devctrl MODEL
A session can end while its workspace is retained. Compute can change while history stays in the controller.
devctrl VS Herdr
Run the session. Manage its environment.
Use devctrl to prepare persistent Claude, Codex and Antigravity workspaces on your servers, with 24 more checked agent CLIs such as GitHub Copilot CLI, Cursor Agent, Qwen Code and Grok Build, then manage sessions, account capacity, file exchange and project access together.
Herdr organizes terminal processes into workspaces, tabs and panes across SSH machines. You supply the host environment; its single binary and community plugins provide the terminal workflow.
With devctrl, create or reuse a managed container, deliver selected credentials and connect the resources the project needs. Update its tools and remove declared task artifacts from the same console.
Choose devctrl to manage both the coding sessions and the infrastructure that keeps them running.
Sources: Herdr model · Herdr SSH · Herdr source · devctrldevctrl VS cmux
Keep the control plane on your servers.
cmux provides SSH workspaces, agent notifications, remote file uploads and supported native session resume. Its Cloud offering adds persistent Linux machines and durable terminals that keep running when the laptop disconnects.
Cloud is in early access and includes Docker, file push/pull, CodeRouter quota views and account failover, with machine-level egress policies.
devctrl lets you operate the controller on your server, create managed project containers and connect selected folders and services. Give developers private environments and budgets across the execution hosts you choose.
Choose devctrl when running the infrastructure yourself is part of the workflow.
Sources: cmux SSH · cmux Cloud · cmux workspaces · cmux files · CodeRouter · cmux security · devctrl · Server usersdevctrl VS Superset
Bring the environment into the workflow.
Superset manages task worktrees, terminals, diffs and review in an IDE. Its terminal daemon survives app restarts; a headless host and remote relay make dedicated machines accessible from desktop and mobile clients.
Account quotas, alternate profiles and cross-agent context handoffs are built in. Setup and teardown scripts can manage Docker.
devctrl prepares and tracks the agent containers, connects project resources and manages user budgets. Use the terminal or browser to run the sessions and maintain the infrastructure behind them.
Choose devctrl to centralize workspace preparation, resource access and server operations alongside your coding sessions.
Sources: Superset model · Superset terminals · Superset hosts · Superset remote · Superset accounts · Superset handoffs · Superset scripts · devctrl · Server usersdevctrl VS Conductor
Choose the hosts. Keep the workflow.
Conductor supports local workspaces on a Mac and isolated Cloud sandboxes. Cloud sessions continue after the desktop app closes; automated file sync connects Cloud workspaces with the local machine. Local sessions have different shutdown behavior.
Cloud also supports collaborative workspaces and API/MCP control. Its documentation says self-hosted Cloud machines are not yet available. devctrl can use existing remote hosts and optional compute adapters while keeping the same console.
Choose devctrl to run persistent workspaces on your own hosts, with the controller, accounts and resources under your management.
Sources: Conductor model · Conductor Cloud · Conductor security · Conductor API · devctrldevctrl VS Happy
Keep your native CLIs. Manage their compute.
Happy supports remote machines and encrypted phone access. Its original CLI wraps native Claude and Codex sessions and has a web client. That CLI is in maintenance mode; new features ship in Desktop.
Newer Desktop uses Happy Harness rather than simply hosting the native CLIs. It adds durable sessions, account pools, cross-model conversation switching, an OS sandbox and configurable Docker. You can self-host its remote runtime.
devctrl runs the native Claude, Codex and Antigravity CLIs, plus 24 more checked agent CLIs such as GitHub Copilot CLI, Cursor Agent, Qwen Code and Grok Build, in managed workspaces on your servers. Connect project resources, move files and manage containers alongside the sessions. Give developers private environments and aggregate compute budgets; share session access explicitly when needed.
Choose devctrl to operate the native coding sessions, their environments and your server infrastructure together.
Sources: Happy CLI · Happy remote · Happy models · Happy permissions · Happy teams · devctrl · Server usersdevctrl VS Emdash
Give remote tasks a managed environment.
Emdash is a free, open-source desktop workspace for native coding agents, task worktrees, diffs and pull requests. Projects can run on your own SSH hosts, including remote file editing and agent installation. Configured tmux keeps real agent and terminal processes alive across app restarts or SSH reconnects.
Its remote-task feature can call your provision and terminate scripts to create an SSH-accessible VM, pod or Docker workspace when enabled in your build. The environment and infrastructure policy remain your responsibility. Shared project scripts and MCP configuration are separate from a server controller with private users and aggregate resource budgets.
Choose devctrl to prepare remote containers, connect project resources and manage sessions, server capacity and cleanup from the terminal or browser.
Sources: Emdash pricing · Emdash tasks · Emdash SSH · Emdash persistence · Emdash provisioning · Emdash settings · devctrl · Server usersdevctrl VS Nimbalyst
Move the work onto your servers.
Nimbalyst runs vendor-native agents inside an open-source visual workspace. It combines searchable sessions, worktrees, tracked file edits, diagrams, mockups and project tools. The iOS companion starts, resumes and steers agents that execute on the paired desktop; session history and window state are stored locally.
Teams share selected documents, trackers and team chat while each teammate runs their own local agents. The browser console edits shared documents. Enterprise can host the collaboration server in your Cloudflare account; that deployment is a collaboration service rather than a documented remote agent-execution controller.
Choose devctrl to move coding sessions, dependencies and generated files to persistent remote workspaces with private environments for each developer.
Sources: Nimbalyst overview · Nimbalyst persistence · Nimbalyst mobile · Nimbalyst Teams · Nimbalyst web · Nimbalyst Enterprise · devctrl · Server usersdevctrl VS Warp
Operate your own agent control plane.
Warp combines a terminal and its own Warp Agent runtime with support for native coding-agent CLIs. Published sessions support browser and mobile viewing and steering. Remote Control keeps a local session in sync while that session remains active; cloud agent runs have a separate execution lifecycle.
The Automation Platform provisions per-run environments, injects scoped secrets and stores transcripts, artifacts and costs. Enterprise self-hosting runs workers on your infrastructure, while Warp operates the control plane. Docker, Kubernetes and direct-host backends have different isolation boundaries. Warp credits meter model, compute and platform usage; native CLI subscriptions keep their provider billing.
Choose devctrl to host the controller yourself and manage native Claude, Codex, Antigravity and other supported agent CLI sessions on the execution infrastructure you choose.
Sources: Warp overview · Warp CLI agents · Warp session sharing · Warp Remote Control · Warp architecture · Warp self-hosting · Warp execution security · Warp billing · devctrl · Server usersdevctrl VS Coder
Infrastructure for your coding-agent sessions.
Coder provisions development environments on infrastructure you control. Terraform templates describe containers, Kubernetes pods and VMs, including their tools and limits. It supports browser terminals, IDEs, workspace sharing, Dev Containers and Coder Desktop file sync.
Current Coder also includes its own agent platform. Coder Agents runs the agent loop in the self-hosted control plane, stores chats outside workspace compute, supports subagents and keeps model API keys out of workspaces. Premium adds resource-credit quotas, AI spend budgets, audit logs and Agent Firewall.
devctrl focuses on native Claude, Codex and Antigravity workflows, with 24 more checked agent CLIs such as GitHub Copilot CLI, Cursor Agent, Qwen Code and Grok Build: persistent sessions, prepared containers, quota-aware account continuation, supported import and task handoffs, and project access to files and services.
Choose devctrl when native Claude, Codex and Antigravity sessions are the work you want to run and maintain on your servers.
Sources: Coder platform · Coder Agents · Coder access · Coder Dev Containers · Coder pricing · Coder quotas · Coder AI budgets · devctrl · Accountsdevctrl VS Daytona
Start the coding session. Keep its context.
Daytona provides sandbox lifecycle, filesystem, process, terminal and computer-use APIs through SDKs and MCP. Containers preserve files across stop/start; Linux and Windows VMs also support memory-preserving pause/resume, hot snapshots and forks. It offers GPU compute and macOS through use.computer. Volumes and snapshots can outlive a sandbox.
Its infrastructure controls include hard sandbox resource limits, network allowlists, organization quotas, HTTPS secret substitution and per-sandbox cost reporting. Resource RBAC governs platform operations; members of a collaborative organization share access to its running sandbox runtimes.
You can use hosted compute or Enterprise BYOC runners. Documented BYOC uses Daytona’s control plane; its old public core repository stopped receiving updates after June 2026. Daytona supplies compute to an agent application, while devctrl supplies a native CLI session workflow with server-user controls.
Choose devctrl for a ready native coding-agent workflow with Claude, Codex, Antigravity and 24 more checked CLIs such as GitHub Copilot CLI and Cursor Agent: prepared environments, retained sessions, account continuation, file exchange and private server users.
Sources: Daytona platform · Daytona sandboxes · Daytona persistence · Daytona isolation · Daytona secrets · Daytona limits · Daytona billing · Daytona organizations · Daytona BYOC · Daytona source · devctrl · Server usersdevctrl VS Cloudroom
Compare the runtime boundary and the operating workflow.
Local and cloud agent workflows with a self-hostable Linux Core. Hosted sandboxes and own-machine Core have different isolation and account boundaries.
Evaluate the documented mode you intend to run. Missing documentation is unknown support, and source review is not a live product test.
Sources: Cloudroom: local/cloud harnesses, Teleport and subagents · Cloudroom Core: self-hosted Linux runtime and account boundary · Cloudroom: durable prompt queues and native conversation recovery · Cloudroom: free local/cloud tier and paid cloud plansdevctrl VS Amika
Choose compute and control-plane ownership separately.
Agent VM workstations on your machines or cloud accounts with private connectivity. Current CLI operations use the hosted control plane; a complete supported self-hosted control plane is not established.
Evaluate the documented mode you intend to run. Missing documentation is unknown support, and source review is not a live product test.
Sources: Amika: own-compute VMs and private networking · Amika: open components, hosted control plane and Rig interfaces · Amika: removed local CLI mode and current self-hosting limitationsdevctrl VS Orca
Compare the runtime boundary and the operating workflow.
Agent IDE with local worktrees, arbitrary SSH hosts, a self-hosted server, CLI orchestration and experimental per-workspace cloud recipes. A worktree alone is not a container or VM boundary.
Evaluate the documented mode you intend to run. Missing documentation is unknown support, and source review is not a live product test.
Sources: Orca: local, SSH, self-hosted server and experimental BYO cloud recipes · Orca: arbitrary SSH hosts, persistent remote terminals and forwarding · Orca: self-hosted servers, revocable pairing and reconnect · Orca: CLI harness support, worktrees, desktop/mobile and MIT sourceFILES + BROWSER
Give sessions context. Get their work back.
Drop a screenshot or dataset into a remote workspace. Download the report, build or archive when it is ready. devctrl handles file exchange through the browser or terminal console, directly with the server.
Keep selected Exchange folders synchronized through the optional SSH Docker client. HTTPS/WSS synchronization is a preview with explicit owner/device bindings; clean restart recovery still has a known limitation. Attach host-local, SSH or S3 folders with read-only or read/write grants for ongoing access.
Herdr supports clipboard images. Its community Mobile Relay adds browser prompts, notifications, approvals, attachments and read-only file views; archive downloads require external tools.
Keep context, generated artifacts and connected storage within the project’s normal workflow.
Sources: File exchange · Folder sync · devctrl folders · Herdr workflow · Mobile RelayRESOURCES + PERMISSIONS
Connect once. Grant access per project.
Make an API, database, model endpoint or folder available to selected projects. Deliver scoped runtime credentials, choose read-only or read/write folder access and reuse method and rate limits on managed MCP routes.
Built-in tools and skills show agents how to use the connections you configure. Review additional access requests, adjust their scope and deliver changes to the running session without restarting it.
Approved operational work uses an isolated read-only task chat. Follow the task, its blocked or completed outcome and a separate notification to the exact requesting session; human chats and messenger routes remain independent.
Connect Telegram, Slack or Teams to inspect output, send prompts and review requests in a paired private bot chat.
Service, folder and MCP gateways enforce their managed access boundaries. Keep upstream database permissions scoped on the infrastructure you operate.
New or explicitly rebuilt restricted workspaces use an isolated Docker network and allowlisted Internet gateway on rootful Docker 28+ or Rootless Docker 29.9.1+. Required Claude/Codex/Antigravity model and sign-in hosts stay available; provider groups and individual API domains use editable private templates. Existing sessions keep their network until a rebuild; older rootless daemons and Boat-native networking are outside this boundary.
Choose access packages at shared system, user, project-group and repository scopes. Team administrators approve shared templates and control additional destinations. Explicit repository policies stay unchanged until you choose inheritance.
The default template covers package registries, toolchain installers, container registries, documentation and many database hosts. Shared object storage and user-content hosts such as Google Cloud Storage, Hugging Face, cloud container registries, JitPack, Read the Docs and Stack Overflow ship as unchecked opt-in groups, because a compromised dependency could use content anyone can upload there for command-and-control or data exfiltration.
Add the Web fonts package to a private Internet-policy template or scoped defaults. It allows Google Fonts stylesheets and font files; saving policy and verified gateway application remain separate.
Give sessions useful resources without repeating the connection setup for every project.
Sources: devctrl · Approvals · MCP policies · devctrl egress · devctrl operator executionACCOUNTS + CONTINUATION
Keep the task moving when an account hits its limit.
See quota and reset timing before starting a session. When an account is exhausted, choose another eligible subscription and continue the task using supported conversation history.
Move work between Claude, Codex and Antigravity while retaining the original session. Use supported Claude/Codex imports or a bounded saved task handoff; Antigravity uses the handoff rather than native transcript conversion.
Choose a hosted or self-hosted Decisions advisor for downgrade-only subagent routing. Review API-equivalent forecasts, project budgets and provider-attributed history.
Choose an account-specific HTTP, SOCKS or existing VPN-gateway route in a supported restricted workspace. Saving a route leaves working sessions on their current connection and applies the change at their next provider process start.
Herdr’s community Agent Router recommends agents and models using subscription and quota information. The table distinguishes those recommendations from continuing an existing conversation.
Set administrator ceilings for agent and Git accounts, then narrow choices per Project or repository. Every linked repository is checked before launch or resume; running sessions keep their existing accounts.
Defaults choose an account; locks define which accounts are allowed. Shared account access and team permissions still apply.
Keep account limits from becoming a lost-context problem: see capacity and choose a supported continuation in the same console.
Sources: Accounts · devctrl account routes · devctrl · Agent Router · Limit account choices for each Project.WORKSPACES + DOCKER
Prepared workspaces. Managed upkeep.
devctrl prepares the repository and toolchain, creates the workspace and delivers scoped runtime credentials. Containers and provider resources have their own lifecycle, so retained session history stays separate from compute.
Each new coding session gets its own working directory and Git branch, even in a shared container. Reuse compatible compute or start a separate environment. Upgrade selected workspaces and review declared task artifacts for cleanup.
Herdr starts agent processes in the environment where its server runs. You can run that server inside a container or sandbox you prepare; Herdr supports wrapper-based agent detection.
Repositories are grouped by Project, with the current Project first. Other Projects and repositories sort by name. Same-organization linking is the default; allow both organizations in Settings before linking across them.
Select optional tools in a template, prepared image or running container. Inspect observed versions, official changelogs and recorded advisory coverage; unknown evidence stays explicit. Component changes preserve active sessions and management connections. Advisory matching remains a preview.
Let devctrl handle the workspace lifecycle, from preparation and updates to reviewing the leftovers.
Sources: devctrl · Herdr CLIs · Herdr model · Choose allowed organizations · Choose optional workspace componentsUSERS + RESOURCE BUDGETS
A shared server with separate workspaces.
Server mode gives each devctrl user private projects, credentials, histories and containers. Administrators publish selected resources and set aggregate compute budgets across a user's containers.
Give users separate checkouts, runtime directories and containers. Choose provisioned storage for new workspaces, track aggregate reservations and measured use, and enable hard CPU, memory and disk caps through verified host adapters. Configure CPU affinity without restarting sessions; selected cores are not exclusive reservations.
An owner can share a running session with named server users: view live output or control prompts and replies. One participant holds input at a time. Sharing leaves runtime ownership, resource grants and account quota with the original owner; it does not merge users’ private projects or containers.
- Shared sessions
- One live conversation can have view-only or control access. Viewing follows output; control additionally lets another person prompt or steer the agent.
- Same repository, separate environments
- Teammates can work on the same Git repository in their own checkouts and containers. devctrl stores private Project records for each owner and separates paths by user; Git carries their code changes. This does not require shared credentials or a shared physical container.
- Multi-user infrastructure
- Application identities own private credentials, execution environments and resource budgets. Selected services may be published without mixing their containers or sessions.
Put shared servers to work with private developer environments and resource budgets in one control plane.
Sources: Server users · devctrl shared sessions · devctrl containersVISIBILITY + EVENTS
See the work. Understand the wait.
Opening a session shows the actual account, access, workspace and terminal stages with elapsed time. Native pending tool calls and reported background work are distinct from questions, permission requests, provider errors and current blocked goals. Unconfirmed long waits stay explicit. Give sessions useful display names without changing native history, and see their active subagent counts when the installed CLI supplies readable lifecycle data.
Inspect native model and reasoning effort observations alongside token counters and reported context-window use for the session and its subagents. Configure ordered native status fields for selected agents; existing custom renderers keep precedence and layouts apply on the next safe start. Optional server-side compact waits for explicitly completed work and a native completed turn, with no pending questions, active children or attached terminal. Configure it globally, per project or per session; it is off by default.
Browse repository compaction history across sessions. Filter and page through confirmed results, native counters and held attempts; missing measurements stay explicit.
Inspect why an empty container waits, stays running or stops. Queued commands, confirmed host outcomes and unconfirmed results remain separate; retained work stays available.
RAM monitoring keeps running on the execution infrastructure after you close the client. Current managed sessions receive measured pressure warnings at safe input points, with the scope and limit that caused them. The monitor does not kill unrelated work or predict the next OOM.
Browse operation events independently of the full project inventory. Download your visible journal as redacted JSONL or connect an authenticated, resumable live collector. These records cover controller operations and access changes, rather than every command the agent runs.
Host cards keep UTC clock differences separate from time-zone offsets, with sample age and uncertainty. A missing physical-host timezone stays unavailable; the display changes no server clock.
Control Room has its own human-attention count. Open the current approval or chat without sending input or losing your draft. Running tasks and completed history stay out of the count.
Request Claude’s native arrival-time markers at the next provider start. Existing settings remain intact, and rendering depends on Anthropic’s rollout. Codex and Antigravity have no verified native toggle.
Matching controller and host versions prepare a session-bound caller for eligible managed Docker sessions on a new provider start or existing broker recovery. Native metadata and guide receipts establish readiness; the running client’s tool list stays separate.
Keep sessions, resource warnings and the operation trail visible without leaving your laptop running.
Sources: devctrl session opening · devctrl session labels · devctrl context telemetry · devctrl auto compact · devctrl RAM warnings · devctrl events · devctrl host clocks · Find requests needing you · Choose native message timestamps · Recover scoped MCP accessTMUX / ZELLIJ + SSH + DOCKER
Build your own workflow, or keep it in one place.
tmux and Zellij already solve detach and reattach for terminal sessions. Zellij also includes an authenticated browser client and installable web app. Add SSH, Docker, file-transfer tools and your own scripts, and you can assemble a capable remote setup.
You maintain the connections between those pieces: which container owns a session, which account it uses, how files arrive, what credentials it receives and what can be deleted later.
devctrl keeps that relationship explicit: session → workspace → compute. Reuse a compatible container or start a separate one, continue on another eligible account, update project tools and review task artifacts for cleanup.
Keep tmux or Zellij when you want to compose the workflow yourself. Choose devctrl when coordinating the pieces is the problem.
Sources: tmux · Zellij · Zellij web · devctrldevctrl VS Paseo
Choose how you manage the execution environment.
Paseo brings agents, files, terminals, diffs and a desktop browser into one workspace. Its daemon runs on your laptop or server, with desktop, mobile, web and CLI clients. Local directories and Git worktrees provide different ways to organize parallel tasks.
It also documents cron schedules for new agents, heartbeats for existing conversations, cross-provider delegation and Hub triggers from GitHub, Slack and Discord.
devctrl manages coding sessions alongside their containers, scoped project resources and private user environments. Paseo offers a Docker image for its daemon, but you install the agent CLIs and configure mounts; that is different from a managed container lifecycle for each workspace.
Choose Paseo for its integrated agent development workflow. Choose devctrl when operating session environments, resource access and self-hosted compute together is the priority.
Sources: Paseo: Local daemon and clients · Paseo: Workspaces and worktrees · Paseo: Schedules and heartbeats · Paseo: Agent orchestration · Paseo: Hub event triggers · Paseo: Docker deployment · devctrl · Server usersdevctrl VS T3 Code
Control your agents from anywhere, or operate their environments.
T3 Code is a free, MIT-licensed control surface for Claude, Codex, Cursor, Grok Build, OpenCode, Antigravity and other provider CLIs that are already set up on a machine. Its server runs on your computer, a server or a VPS; desktop, web and mobile clients connect through T3 Connect, the local network, Tailscale or SSH.
Threads use their own branches or Git worktrees, with diff review, pull requests, pooled subscription limits, provider switching with context handoffs, scheduled and webhook tasks, and an MCP server for outside agents.
devctrl manages coding sessions alongside their containers, scoped project resources and private user environments. T3 Code runs agents as processes on the host it is installed on; containers, per-user isolation and resource budgets are not part of its reviewed documentation.
Choose T3 Code for a polished multi-provider client on machines you already prepared. Choose devctrl when operating session environments, resource access and self-hosted compute together is the priority.
Sources: T3 Code: Agent harness control surface · T3 Code: Pricing · T3 Code: Remote access · T3 Code: Project settings · T3 Code: Usage history and pooled subscription limits · T3 Code: Context handoffs between providers · devctrl · Server usersdevctrl VS Anvaya
Run your own harness, or manage the native CLIs.
Anvaya is a coding agent harness rather than a session controller. It runs its own agents against local Ollama models or bring-your-own-key Anthropic and OpenAI-compatible endpoints, and keeps what each project teaches in a local memory graph that other clients can query over MCP.
It documents sandbox profiles, subagents and supervised swarms in shared, worktree or cloned workspaces on one Linux or macOS machine. It is pre-1.0: source access is by waitlist and pricing is not set.
devctrl runs the native Claude, Codex and Antigravity CLIs in managed containers on your servers, with project resources, private user environments and a web console. Anvaya’s reviewed documentation does not describe remote servers, containers or multi-user operation.
Evaluate Anvaya as a local harness and memory layer. Choose devctrl to operate native agent sessions, their environments and shared server infrastructure.
Sources: Anvaya: Agentic coding harness with an experience store · Anvaya: Concurrency · Anvaya: Private source access · Anvaya: Providers and models · Anvaya: Sandbox and permissions · Anvaya: Agent swarms · devctrl · Server usersdevctrl VS boxes.dev
Rent a cloud computer per agent, or run agents on your own servers.
boxes.dev gives every Claude or Codex agent its own cloud Linux machine. Setup agents port your local dev environment to a Template box, and each new devbox starts from its snapshot with separate files, processes, ports and database state.
You steer agents from desktop, mobile, web or the dvb CLI, and start them from Slack, Linear, schedules or an API call. It uses your own model subscription or API key and bills awake machine time per box-hour; idle machines sleep for free.
devctrl runs the same native CLIs in managed containers on servers you own, with private user environments, project resources and an operator Control Room. boxes.dev’s reviewed documentation describes vendor-hosted machines; BYO cloud is listed only for its team plan.
Choose boxes.dev for vendor-hosted machines per agent without running servers. Choose devctrl to run native agent sessions on your own servers with shared resources and access control.
Sources: boxes.dev: Cloud computers for Claude and Codex · boxes.dev: How boxes.dev works · boxes.dev: Connect Codex and Claude · boxes.dev: Your data and credentials · boxes.dev: Sleep · devctrl · Server usersPLAN YOUR SETUP
Your infrastructure. Your choices.
- Choose your compute.
- Use existing servers, SSH hosts, local Docker or optional sandbox providers. You operate the control plane and choose where projects execute.
- Keep access available.
- Host the controller and browser gateway on a server to use the console while your laptop is offline.
- Continue with the context you have.
- Account and provider transfers use supported native import or a saved task handoff. The original session remains available; history support depends on the chosen clients.
- Set budgets that match the host.
- Give users separate containers and aggregate resource budgets. Enable CPU, memory and storage hard caps through verified host adapters.
- Keep networking explicit.
- Restricted workspace networking requires a new/rebuilt runtime on rootful Docker 28+ or Rootless Docker 29.9.1+; model and sign-in hosts stay allowed. Older rootless daemons and Boat-native networking are outside this Docker boundary.
- Publish the console your way.
- Save a tunnel provider or your HTTPS origin for the next web start. Free tunnel URLs can change; ngrok needs its agent and credentials, and an own domain needs your DNS, TLS and reverse proxy.
- Connect only what you use.
- Add compute, messenger and Decisions providers as optional packages. Model advice requires configuration; cost forecasts estimate API-equivalent usage rather than subscription savings.
PRIMARY SOURCES
Go straight to the documentation.
Follow the product and extension documentation behind the selected comparison.
- devctrl: managed RTK OSS, sgrep, ast-grep, Repomix and Graphify, with live installation and Git-aware graph refresh
- devctrl: current account model catalogs, defaults and optional Antigravity quota fallback
- devctrl: personal account display order shared by TUI, web and desktop
- devctrl: frozen session snapshots, native resume and encrypted S3 recovery
- devctrl: opt-in idle cloud shutdown, encrypted S3 archives and restore
- devctrl: shipped features and limitations
- devctrl: direct browser and terminal file transfers
- devctrl: optional prebuilt local synchronization client
- devctrl: account and provider continuation
- devctrl: Claude reset offers and Codex reset credits — checked
- devctrl: MCP access requests and delivery
- devctrl: managed MCP method and rate policies
- devctrl: private server users and resource budgets
- devctrl: Decisions and chat integrations
- devctrl: opt-in Public Project MCP for external OAuth clients
- devctrl: Task Graph and the Projects Activity view
- devctrl: Project messaging between agent sessions
- devctrl: OpenRouter, Vercel AI Gateway and LiteLLM accounts for Claude
- Herdr: workspaces, tabs and panes
- Herdr: supported agents and sandbox wrappers
- Herdr: agent management skill
- Herdr: SSH machines and persistent sessions
- Herdr: agent workflow and clipboard images
- Herdr Mobile Relay: browser access, notifications and files
- Herdr Agent Router: quota-aware agent and model routing
- Herdr: implementation, platforms and supported agents
- cmux: SSH workspaces and remote session persistence
- cmux: persistent Cloud machines in early access
- cmux: durable Cloud terminals and installed tools
- cmux: Cloud uploads, downloads and secret files
- cmux: CodeRouter quotas and subscription failover
- cmux: Cloud isolation and machine network policies
- cmux: agent resume and process restoration boundaries
- cmux: terminal, notifications, CLI and skills
- Superset: worktree-first IDE and task workspaces
- Superset: persistent terminals survive app restarts
- Superset: dedicated hosts, remote relay and teammate access
- Superset: headless host daemon and lifecycle
- Superset: account quota, profiles and shared Claude history
- Superset: native session forks and cross-agent handoffs
- Superset: file explorer and editor
- Superset: running and waiting agent notifications
- Superset: MCP workspace and agent management
- Superset: setup and teardown scripts including Docker
- Superset: Slack task and coding-agent dispatch
- Conductor: parallel sessions, workspaces and review
- Conductor: Cloud persistence, hosting and account access
- Conductor: cloud/local file sync and permission boundaries
- Conductor: workspace API and MCP management
- Conductor: shared chats and separate workspaces
- Conductor: Codex quota and reset display
- Conductor: workspace setup and archive scripts
- tmux: persistent terminal sessions
- Zellij: terminal session restoration
- Zellij: authenticated browser client and PWA
- Happy: original CLI in maintenance mode, native agent sessions and browser access
- Happy: self-hosted remote runtime and server lifecycle
- Happy: durable sessions, subagents and working/waiting state
- Happy: subscription pools, quotas and mid-session model switching
- Happy: encrypted images, file reading and mobile permissions
- Happy: worktrees, setup, file copying and archive cleanup
- Happy: OS sandbox, network restrictions, secrets and MCP review
- Happy: machine/project configuration and Docker images
- Happy: workspace MCP servers and skills
- Happy: team identities and shared projects/sessions
- devctrl: combined repository launches, organization imports and Control Room suggestions
- devctrl: GitHub/GitLab setup, host choices and isolated credentials
- devctrl: controller, runtime, grants and trust boundaries
- devctrl: process, history and storage persistence
- devctrl: installation targets and devcontainer preparation
- devctrl: TypeScript/Node implementation and package manifests
- devctrl: 14 interface languages across shared clients and native onboarding sources
- devctrl: optional source-built Tauri desktop client
- devctrl: Android ARM64 development preview and verification limits
- devctrl: same-owner container reuse and tenant isolation
- devctrl: owner-granted live session viewing, control and input leases
- devctrl: optional sandbox lifetimes, secret safety and accounting
- devctrl: optional sandbox lifetimes, secret safety and accounting
- devctrl: folder access and enforced gateway boundaries
- devctrl: isolated Docker networking and workspace allowlists
- devctrl: paginated operation/access history, redacted JSONL export and live collectors
- devctrl: measured RAM pressure and safe session warnings
- devctrl: server-side CPU, RAM and disk I/O history
- devctrl: account-scoped HTTP, SOCKS and existing VPN gateway routes
- devctrl: session display names and active native subagent counts
- devctrl: one Main administrator conversation, native conversation history, local system-voice reply playback and audited server MCP operations
- devctrl: observed host UTC samples, physical timezones and uncertainty
- devctrl: approval policies and owner-scoped remembered task intents
- devctrl: independent main-session Decisions policy and native acknowledgement
- devctrl: control-plane update progress and data-preserving recovery
- devctrl: canonical operator host commands and typed, tracked diagnostics or managed-service repair
- devctrl: finite current-container resource requests and verified limit application
- devctrl: weekly native account attribution and limits
- devctrl: project activity, native token history and Git totals
- devctrl: native session and subagent token/context observations
- devctrl: scheduled native progress questions and timestamped session headers
- devctrl: opt-in compact after explicitly completed work
- devctrl: saved web access providers and public URL prerequisites
- devctrl: verified storage quotas, user workspace placement and CPU affinity
- devctrl: per-session browser previews on an isolated origin
- devctrl: risk-based screening of text that agents send each other
- devctrl: session opening stages and terminal attachment
- Herdr: free Apache-2.0 runtime, platforms and machine restart
- Herdr: Rust implementation and dependencies
- cmux: Free/Pro/Max/Team and Enterprise self-hosted compute
- cmux: machine limits, snapshots, recovery and scoped tool updates
- cmux: cross-platform TUI client
- cmux: Swift/AppKit client source and Ghostty integration
- Superset: Free/Pro pricing and Enterprise identity/audit
- Superset: supported desktop/mobile/host platforms
- Superset: TypeScript/React/Electron implementation
- Superset: organization projects, worktrees and Cloud clones
- Conductor: free local usage and paid Cloud/team plans
- Conductor: cloud microVMs, iOS and live collaborative workspaces
- Conductor: bundled native CLIs and inference billing
- Happy: desktop platforms, MIT licensing and provider subscriptions
- Happy: daemon state, durable replay and encrypted clients
- Happy: TypeScript runtime and package manifests
- Emdash: desktop workflow, platforms and native CLI agents
- Emdash: source, license, local state and provider hooks
- Emdash: Electron desktop implementation manifest
- Emdash: desktop-managed remote daemon, pinned artifacts and SSH socket transport
- Emdash: free Apache-2.0 desktop
- Emdash: task worktrees, model picker and existing-directory tasks
- Emdash: provider detection, installation, status hooks and resume
- Emdash: SSH execution, SFTP files and remote installation
- Emdash: enabled-build remote tasks using provision and terminate scripts
- Emdash: optional local and remote tmux process persistence
- Emdash: shared project settings, environment-file copies and lifecycle scripts
- Emdash: MCP configuration synced to native agents
- Emdash: native Claude and Codex quota meters and reset times
- Emdash: contact-based Cloud workspaces
- Nimbalyst: native agents, visual editors, worktrees and collaboration boundary
- Nimbalyst: MIT desktop and SwiftUI iOS source
- Nimbalyst: Electron, React and editor implementation
- Nimbalyst: Individual, Teams beta and Enterprise pricing
- Nimbalyst: searchable sessions and background session launch
- Nimbalyst: SQLite/PGLite session history, window state and backups
- Nimbalyst: local per-project token and tool activity reports
- Nimbalyst: Codex native account limits and reset tracking
- Nimbalyst: Claude native subscription limit tracking
- Nimbalyst: trust gate, project approvals, directories and URL patterns
- Nimbalyst: app/project MCP servers, OAuth and native workspace tools
- Nimbalyst: iOS control of desktop agents and optional alpha markdown sync
- Nimbalyst: selected document and tracker sync; each teammate runs local agents
- Nimbalyst: shared-document browser console and organization roles
- Nimbalyst: Enterprise collaboration server in your Cloudflare account
- Nimbalyst: experimental native agent teams and iterative Super Loops
- Warp: terminal, Warp Agent and Automation Platform execution models
- Warp: open-source client and MIT/AGPL license boundaries
- Warp: Rust workspace and native UI crates
- Warp: current free, paid and Enterprise plans and credits
- Warp: model, compute and platform credit accounting
- Warp: native CLI agents in the terminal
- Warp: browser and mobile control of published native CLI sessions
- Warp: shared live sessions, access roles and steering
- Warp: hosted control plane and hosted or Enterprise self-hosted execution
- Warp: Enterprise managed workers and unmanaged CLI execution
- Warp: existing CI or hosts invoke the Oz CLI
- Warp: Docker images, repository setup and runtime environment
- Warp: Docker, Kubernetes and direct-host isolation and networking
- Warp: action approvals and command/MCP server policies
- Warp: model selection, Auto routing and provider-capacity fallback
- Warp: user/team billing, credit pools and integration identities
- Warp: Enterprise organization roles and enforced policies
- Warp: runtime secret allowlists and user/team secret scope
- Warp: operating system, container/VM and CPU/memory runner configuration
- Warp: local/cloud handoff and workspace snapshots
- Coder: self-hosted workspaces, Terraform infrastructure and native Coder Agents
- Coder: open-source implementation, server binary and local Docker quickstart
- Coder: Community, Premium and AI Premium plans
- Coder Agents: control-plane agent loop, persistent chats, subagents and downloadable attachments
- Coder Agents: organization model permissions, providers and optional personal API keys
- Coder Agents: enforced administrator configuration and template permissions
- Coder Agents: external MCP servers, authentication and tool allow/deny lists
- Coder: persistent web terminals, SSH, web IDEs and Coder Desktop bidirectional file sync
- Coder: template-installed browser file uploads and downloads
- Coder: user and group workspace sharing with use/admin roles
- Coder Agents: read-only conversation sharing and live updates
- Coder: user identities, groups, roles and Premium custom roles
- Coder Premium: resource-credit budgets enforced across a user’s workspaces
- Coder Premium: approximate AI Gateway spend, user/group budgets and spend exports
- Coder Premium: process-level Agent Firewall, network allowlists and centralized HTTP audit logs
- Coder: Prometheus metrics, logs and workspace observability
- Coder: devcontainer.json integration through Docker and the Dev Containers CLI
- Daytona: sandbox infrastructure accessed through SDKs, API and CLI
- Daytona: public core repository frozen after moving development to a private codebase in June 2026
- Daytona: per-second compute pricing and Enterprise features
- Daytona: control plane, sandbox runners, toolbox and persistent volumes
- Daytona: containers, Linux/Windows VMs, GPU sandboxes and lifecycle policies
- Daytona: filesystem persistence, VM memory persistence and independent volumes
- Daytona: runtime hard limits, filesystem isolation and sandbox network boundaries
- Daytona: streaming SDK/API file uploads, downloads and filesystem permissions
- Daytona: configured external storage mounts alongside shared volumes
- Daytona: authenticated browser terminal for running sandboxes
- Daytona: MCP sandbox management, files, code execution and computer-use tools
- Daytona: personal/collaborative organizations, role assignments and organization-wide runtime access
- Daytona: organization compute pools, per-sandbox resources and API rate limits
- Daytona: encrypted secrets substituted into outbound HTTPS headers, with configured host allowlists
- Daytona: organization user and sandbox lifecycle audit logs
- Daytona: organization wallet, resource usage and per-sandbox cost attribution
- Daytona Enterprise: your runner infrastructure with Daytona’s control plane
- Cloudroom Core: self-hosted Linux runtime and account boundary — checked
- Cloudroom: local/cloud harnesses, Teleport and subagents — checked
- Cloudroom: durable prompt queues and native conversation recovery — checked
- Cloudroom: free local/cloud tier and paid cloud plans — checked
- Amika: own-compute VMs and private networking — checked
- Amika: open components, hosted control plane and Rig interfaces — checked
- Amika: removed local CLI mode and current self-hosting limitations — checked
- Orca: local, SSH, self-hosted server and experimental BYO cloud recipes — checked
- Orca: arbitrary SSH hosts, persistent remote terminals and forwarding — checked
- Orca: self-hosted servers, revocable pairing and reconnect — checked
- Orca: tracked CLI orchestration and parent/child tasks — checked
- Orca: skill registry and MCP — checked
- Orca: provider usage, reset observations and explicit unavailable values — checked
- Orca: CLI harness support, worktrees, desktop/mobile and MIT source — checked
- Orca: Jira issue drawer — checked
- devctrl: bounded technology profiles, diagnostic checks and catalog limitations — checked
- devctrl: curated manifest-based skill selections, inheritance and explicit pinned installation — checked
- devctrl: review contracts, evidence and server gate boundaries — checked
- Nimbalyst: Compact action and compaction failure recovery
- Warp: manual compact and compact-and conversation commands
- Coder Agents: token-threshold and on-demand context compaction
- Limit account choices for each Project.
- Choose allowed organizations
- Choose optional workspace components
- Find requests needing you
- Choose native message timestamps
- Recover scoped MCP access
- Paseo: Getting started — checked
- Paseo: Local daemon and clients — checked
- Paseo: Remote connections — checked
- Paseo: Security boundaries — checked
- Paseo: Docker deployment — checked
- Paseo: Browser client — checked
- Paseo: Workspaces and worktrees — checked
- Paseo: Supported providers — checked
- Paseo: Codex authentication and billing — checked
- Paseo: CLI and account quota reports — checked
- Paseo: Agent orchestration — checked
- Paseo: Agent profiles — checked
- Paseo: Schedules and heartbeats — checked
- Paseo: Hub event triggers — checked
- T3 Code: Bring your own subscription, per-thread branches and pull requests — checked
- T3 Code: Agent harness control surface, providers and installation — checked
- T3 Code: Pricing, remote use, local models and data handling — checked
- T3 Code: Remote access, SSH environments, scopes and linked environments — checked
- T3 Code: Background service and updates — checked
- T3 Code: Usage history and pooled subscription limits — checked
- T3 Code: Project settings, worktrees, cleanup and scheduled tasks — checked
- T3 Code: Context handoffs between providers — checked
- T3 Code: Mobile notifications for approvals and input — checked
- T3 Code: Message attachments and queued messages — checked
- T3 Code: Permission modes and provider enforcement — checked
- T3 Code: Machine-wide managed policy — checked
- T3 Code: MCP server for outside agents — checked
- T3 Code: Guest links for one shared thread — checked
- T3 Code: Secret references and agent secret requests — checked
- T3 Code: Codex accounts and usage-limit messages — checked
- T3 Code: Claude accounts, compaction and usage limits — checked
- Anvaya: Agentic coding harness with an experience store — checked
- Anvaya: Private source access, build and first session — checked
- Anvaya: Concurrency, data locality, platforms and pricing — checked
- Anvaya: Installation and requirements — checked
- Anvaya: Providers and models — checked
- Anvaya: Mind MCP memory server — checked
- Anvaya: Sandbox and permissions — checked
- Anvaya: Agent swarms, placement and verification — checked
- Anvaya: Subagents and parallel work — checked
- Anvaya: CLI modes, tools and approvals — checked
- boxes.dev: Cloud computers for Claude and Codex, pricing and FAQ — checked
- boxes.dev: How boxes.dev works — checked
- boxes.dev: Connect Codex and Claude — checked
- boxes.dev: Models, reasoning and usage — checked
- boxes.dev: Agent permissions — checked
- boxes.dev: Your data and credentials — checked
- boxes.dev: Add MCP servers — checked
- boxes.dev: Sleep, wake and recover a devbox — checked
- boxes.dev: Install and use the CLI — checked
- boxes.dev: Security and credential guidance — checked
Put your server to work.
Give your agents persistent remote workspaces. Start with one repository and server, then add projects, developers and resources from the same console.