Voice input
Control Room can turn a short microphone recording into a reviewed prompt. The same React flow runs in the terminal console, browser and desktop client. It never sends a transcript to an agent automatically: record, stop, review or edit, choose Use transcript, then send the prompt separately.
Inside the Control Room chat, Voice input opens inline in the composer area. History stays mounted, and any typed draft is retained. Recording, transcription and transcript review happen there; Use transcript appends to the draft, without sending it. Esc discards the recording panel and returns to that draft, instead of leaving the chat.
Subscription voice
Control Room → Voice input uses subscription mode by default. An available Claude account supplies native CLI dictation without an additional transcription API key. Choose its account in the recording panel if several are available. The source list is scoped to your managed accounts and running workspaces; it reports account authorization and execution-host availability. CLI compatibility is checked when transcription starts; an incompatible runtime is reported then, not presented as a completed preflight check. The managing conversation can use either agent: a Codex Control Room can use an eligible Claude account for voice.
The client records audio, normalizes it to 16 kHz mono 16-bit PCM WAV, and sends it through the controller to a bounded native CLI voice bridge on that account's execution host. A short-lived official Claude process handles only dictation in a private temporary context. It uses an empty private working directory, with no tools, MCP configuration or lifecycle hooks from the working conversation. The bridge exports the unsubmitted composer through the CLI's external-editor action, without pressing Enter or creating a model task. The active session is not restarted or interrupted. The controller pins the source account/workspace and rejects stale or changed sources instead of routing audio elsewhere.
devctrl does not currently provide Codex subscription dictation: the official
audio RPC available for this integration requires API-key authentication. The
interface shows that limitation rather than claiming support for a Codex
subscription. There is no automatic fallback to a paid external transcription service.
The bridge supports the official Linux Claude 2.1.x CLI, with a minimum of
2.1.284, an already signed-in eligible account in a running workspace, and a
host agent reporting native-voice-v1. A different major/minor CLI line needs
its own compatibility check. Its virtual recorder needs no physical ALSA device;
a container with real ALSA sound cards is rejected to avoid selecting unrelated
recording hardware.
Native dictation was verified against Claude 2.1.284 and an existing subscription using synthetic speech; the draft matched and temporary material was removed. Normal tests mock provider traffic and never require paid credentials or microphone hardware.
The selected account must have a valid access token for the complete operation. The helper receives only a temporary copy of that access token and necessary account metadata. It never receives the account's refresh token or changes the shared login store. Near-expiry authorization needs renewal through the normal account workflow before another recording. Audio and temporary credential files are private to the helper and removed when it finishes.
Quota, language and failures
Claude dictation does not consume the model messages or tokens counted by
/usage. An exhausted conversation account can therefore block sending a prompt
without proving that dictation is unavailable. Control Room shows the current
chat account's quota separately. Choose account for a new chat opens the
existing account/host picker, followed by an explicit Start new chat action.
It retains the original chat and its current client draft; it does not silently
change that conversation's account. A failed start with a known new chat ID
retries that chat rather than creating duplicates.
Subscription settings expose the supported spoken language. A saved language is
passed only to the dictation helper's private temporary profile; it does not
rewrite the account's normal Claude language settings. Blank uses the native
default, English. Supported dictation languages are separate from interface
translations. An explicit language and enhanced native error reporting require
an updated host advertising native-voice-diagnostics-v2; old hosts can continue
default-language requests. Update the management host agent for these features,
without restarting working provider sessions.
The browser checks recording signatures before decoding or uploading, and the native bridge validates its PCM WAV contract. Unsupported formats, unavailable services, authorization/policy failures, reported quota failures, missing audio and explicit no-speech results remain distinct. An empty composer without a reported cause returns an unavailable-transcription error, rather than claiming that speech was not recognized. Audio, transcripts and raw provider panes are not included in error responses or diagnostics.
These distinctions follow the official Claude dictation documentation. The diagnostic and interface regressions are verified offline; they do not establish a new live transcription result or microphone-hardware check.
Optional external transcription
Open Control Room → Voice settings and explicitly select External transcription (advanced) only when you want a separate service. External transcription is initially disabled. Its settings are:
- An OpenAI-compatible transcription URL ending in
/audio/transcriptions. - A model offered by that service; models are not assumed to be interchangeable.
- The service's API key when required. A self-hosted service can explicitly disable the credential requirement.
- An optional two-letter language code; blank uses service language detection.
Use HTTPS, or controller-loopback HTTP for a local service such as
http://127.0.0.1:9000/v1/audio/transcriptions. These are controller-side
addresses, not addresses on the device displaying the interface. The service
must accept multipart file, model, optional language, and
response_format=json, and return a JSON text string.
Settings are durable controller integration state. Keys use the existing AES-GCM credential store and are never returned by the settings API or placed in images, session metadata, snapshots or audit messages. Changing the endpoint removes its previous key unless a replacement is supplied. All configuration and transcription routes require Control Room operator authorization.
Microphone location
| Interface | Recording device | Requirements |
|---|---|---|
| Browser | Device displaying the browser | HTTPS or localhost, MediaRecorder and explicit microphone permission |
| Desktop | Device running the desktop application | WebView microphone support and operating-system permission; the macOS bundle includes its microphone usage explanation |
./devctl tui in Docker |
Your local computer | Its existing host-action bridge runs SoX rec, or ALSA arecord on Linux, with a working default input device |
devctrl tui remote |
Your local computer | The standalone Python launcher starts a private microphone bridge and forwards it through the active SSH connection; install SoX or ALSA locally |
| Directly executing the TUI locally | Your local computer | SoX rec, or ALSA arecord on Linux, with a working default input device |
| Bare SSH/container without the launcher bridge | Unavailable | Reconnect with devctrl tui remote, or use the browser or desktop client; devctrl never records from the remote server's microphone |
The interface reports missing recorders, unsupported WebViews and refused microphone access. It does not silently install packages, record in the background or bypass operating-system permission. Camera permission is not requested. The native launcher bridge does not open the microphone until the user starts recording. Its ephemeral bearer is separate from its URL, never saved to connection metadata, and accepted only for its bounded local microphone/speech operations. It is delivered through encrypted SSH standard input into a private controller file (directory mode 0700, file mode 0600); process arguments contain only its path. The launcher removes this file when the TUI exits. If a crashed client cannot complete remote cleanup, the old bearer becomes useless when its local bridge shuts down. Bootstrap failure leaves normal interactive SSH input and TUI startup intact. The bridge binds to loopback, rejects browser-origin requests, and stops when the launcher exits; SSH remote forwarding must be permitted on the server.
Install a missing local recorder
The recording panel probes the actual local microphone device without opening it.
When its launcher detects a supported package manager, it offers the exact local
installation plan, such as Install local recorder with brew install sox on
macOS. Installation begins only after that action (i in TUI). The local launcher
validates the advertised installer ID and runs a fixed command; the controller
and execution host never receive an arbitrary shell command or install audio
packages. Linux plans use the detected native package manager and noninteractive
privilege checks. Interactive sudo passwords are not collected. Unsupported
platforms/managers show an explanation instead of a fabricated install action.
Installation progress, elapsed time and errors stay in the recording panel.
It continues on the local computer if you close the panel; reopening probes its
saved state. Check microphone (r) refreshes status without installing or
recording. A missing local launcher bridge cannot fall back to server installation.
Browser/desktop recording uses getUserMedia, with browser/site and OS permission
help; it does not ask users to install SoX or ALSA.
Limits and privacy
Recordings stop after 20 seconds. Uploads are capped at 700,000 audio bytes, with at most two simultaneous controller transcriptions. External requests have a 25-second provider timeout; native subscription requests have a bounded 85-second bridge deadline and a 95-second client/proxy deadline. These limits keep requests below the existing web gateway body limit. Native recording uses a private temporary directory that is removed on stop, cancellation or error. Docker host-action responses use private, short-lived IPC files removed after consumption or a bounded expiry. Browser tracks stop before audio conversion and when recording is cancelled. Native dictation uses bounded private temporary audio material removed by its bridge. The controller does not persist recordings or transcription responses; a transcript becomes part of the normal native conversation only after you explicitly send its reviewed text.
Audio is sent only after the user starts recording, to the explicitly selected subscription account or external service. Subscription provider terms still apply. An external service can charge separately and apply its own retention policy. Closing the review discards its draft; an already submitted transcription request can still complete. Audit events record settings changes and completed transcription operations, without audio, transcripts or credentials. HTTP redirects are rejected so provider credentials cannot be forwarded to a redirected origin.
This is short-recording dictation, not a continuous voice call or automatic
voice-command execution. Actual microphone hardware and paid transcription
providers are not exercised by the normal test suite. The opt-in native test in
packages/runtime/src/native-voice.integration.test.ts needs a synthetic raw PCM
fixture and an existing authorized account inside the selected Linux workspace.
It is skipped unless both DEVCTL_NATIVE_VOICE_TEST_ACCOUNT_PATH and
DEVCTL_NATIVE_VOICE_TEST_PCM are set; neither value is a credential to publish.
API
GET /v1/control-room/voice: redacted settings, selected mode, owned native sources and availability reasons, selected source ID, and recording limits.PUT /v1/control-room/voice/settings:mode: "subscription" | "external", external settings and optional write-onlycredentialorremoveCredential.POST /v1/control-room/voice/transcribe:{ "mode": "subscription", "sourceId": "opaque-owned-source", "mimeType": "audio/wav", "audioBase64": "..." }; external mode omitssourceId. Returns{ "text": "..." }, plus the native provider/source label where available. The request cannot change its configured mode or inject an endpoint.
All responses use Cache-Control: no-store. The upload cannot override its provider endpoint, authorization header or session destination.
Sources: OpenAI file transcription API, ALSA recording implementation, Tauri macOS privacy configuration.