For the complete documentation index, see llms.txt. This page is also available as Markdown.

Voice Mode

Talk to your coding agent and have it talk back, on desktop and mobile. Voice Mode is in alpha while we polish it.

Turn it on

Open Settings → Voice Mode and flip the toggle. There is no separate release channel to switch. The panel also shows microphone permission status, with a deep link to macOS System Settings if you need to grant access.

The activity bar mic button highlighted while Voice Mode is active
The mic button in the activity bar lights up while Voice Mode is active.

What the voice agent can do

  • Start a new coding session by voice. Say something like "create a new session" and the agent creates one and makes it active, so any follow-up such as "ask the coding agent to..." or "run this prompt" targets the new session without you navigating there.

  • Run "Commit with AI" by voice. The agent surfaces the same commit-proposal widget you get from the UI. Approve or reject by voice.

  • Generate a project summary without an Anthropic API key. "Generate Project Summary" launches a new session in your configured agent (Claude Code, Codex, and so on) rather than calling the Anthropic API directly.

On mobile

The voice agent runs in the mobile app too, and can drive your desktop sessions hands-free:

  • Start a new coding session on your desktop.

  • Find a session by topic, spoken in your own words.

  • Switch between sessions.

  • Summarize a session, including any question it is waiting on.

  • Answer a session's pending question by voice.

  • Send a new coding task to your desktop agent.

The floating mic

While the voice agent is working, a floating mic shows what it is doing at each step. Pause stops it listening or speaking, and Cancel ends the turn, so you always stay in control.

Settings

Voice Mode has its own settings panel under Settings → Voice Mode:

  • Realtime Model. gpt-realtime-2 is the newer model, with stronger reasoning, a larger context window, and more consistent voice. gpt-realtime is the fallback for accounts without access to it.

  • Reasoning Effort. Choose from minimal, low, medium, high, and extra high. Higher is smarter but slower and more expensive; low is a good default for a responsive voice relay. This applies to gpt-realtime-2.

  • Voice. Pick the spoken voice, with a Preview button to hear a sample.

  • Turn detection. Tune how the agent decides you are done speaking: input mode, voice-detection sensitivity, pause before processing, whether it can be interrupted, and how long the mic keeps listening (the listen window, 5 to 30 seconds).

  • Command submission. Set a review delay before a voice command is sent to the coding agent, from immediate up to 10 seconds, so you can catch and edit a misheard request.

  • Language. The voice agent always speaks in your configured preferred language, whatever language you speak to it in.

The Voice Mode settings panel showing turn detection, command submission, project summary, and pricing
The Voice Mode settings panel, under Settings → Voice Mode.

The panel also shows a Project Summary (an AI-generated overview of your project that the voice agent uses for context, which you can regenerate) and OpenAI's usage pricing for voice, since Voice Mode runs on OpenAI's realtime voice models.

The listen window

On desktop, the listen window (15 seconds by default) starts when audio playback drains in your speakers, not when the server finishes streaming, so the mic stays open through the end of the agent's reply. Turns that only call a function also wake the mic, with a short readiness cue when there is no audible reply.

A note on stability

Voice Mode is alpha and may change, break, or be removed without notice. If you hit a problem, share feedback via Discord or our support channels. See Feedback, Discord, Support, Releases.

Last updated