What
Hold the Space bar to dictate text via speech-to-text, with real-time transcription streamed over WebSocket to Claude.ai's /api/ws/speech_to_text/voice_stream endpoint.
Usage
/voice # Toggle voice mode on/offThen hold Space to record; release to stop. The transcript is injected into the input field.
Details
- Requires a Claude.ai account (OAuth); run
/loginfirst if not signed in - Persisted via the
voiceEnabledsetting insettings.json - Audio capture uses a bundled native module (
audio-capture.node) on macOS/Linux, with fallback tosox(reccommand) orarecordon Linux - Offers guided SoX installation (detects
brew,apt-get,dnf,pacman) if no audio tool is found - Supports a "focus mode" that continuously records when the terminal has focus, flushing each finalized segment immediately
- Cursor displays an animated audio-level indicator (bar graph characters
▁▂▃▄▅▆▇█) with color cycling while recording - Shows "listening…" while recording and an animated "Voice: processing…" while finalizing
- Supports 11 languages: English, Spanish, French, Japanese, German, Portuguese, Italian, Korean, Chinese, Hindi, Indonesian — configured via the
languagesetting - Microphone permission checks with platform-specific guidance (macOS: System Settings → Privacy & Security → Microphone; Linux: system audio settings; Windows: Settings → Privacy → Microphone)
- Not available in remote/headless environments — prints helpful message directing users to run locally
Evidence
/voice command (search for "Toggle voice mode"), WebSocket client (search for "/api/ws/speech_to_text/voice_stream"), settings definition (search for "voiceEnabled")
[Gradual Rollout] — Gated behind the tengu_amber_quartz feature flag (defaults to !1/false). The /voice command is hidden when the flag is off. Users with the flag enabled who are logged in to Claude.ai can use this feature.
Strings lifted out of the shipped bundle, so the claim above can be checked against them.