Voice

A box is a directory of your data, kept under version control with a full history of changes (using git); a card is a markdown file with a structured header; the agent is the coding agent (Claude Code or Codex) that operates the box. Voice is a first-class way in and out of the box, on both desktop and phone, not a bolt-on to typing.

What it does for you

• Dictate a chat message with spoken controls: say "send message," "erase message," or "microphone off" instead of reaching for the screen. "Send checkpoint" sends what you have so far and tells the agent the message is partial, so it does not answer yet and the next message continues it.
• Keeps dictating through a brief network drop or the microphone getting grabbed by something else, instead of losing what you said.
• Closes the browser's microphone on its own after five minutes with no new spoken words, keeps the text, and shows a notice saying why, so a forgotten recorder does not stay open.
• Every dictated message gets a second, higher-quality transcription from the box when a transcription key is available; there is no setting for it. The live text on screen is sent only when that pass cannot run or fails, and it is marked as live text so the agent knows. If live transcription drops partway through, the live text shows […] where words are missing until the better transcript fills them in.
• A narration mode for long, loose voice dumps: the box stays quiet, transcribes, and files what you said rather than chatting back line by line.
• A voice memo recorded on the phone becomes an audio card with a transcript and a short summary once transcription finishes.
• A capture session groups a burst of photos and spoken remarks from one sitting into a single readable timeline.
• The agent's replies can be spoken aloud, and the agent chooses the voice, pacing, and emphasis for each spoken reply, not just one fixed narrator.
• Short sounds (earcons) mark recording start, stop, failure, and send, so you know the state of the microphone without looking.
• Siri and Shortcuts can save a short thought or start a voice note without opening the app (phone-capture.md).
• On the phone, speech recognition runs on the device itself. On iOS 26 the high-quality pass over the recording can run there too, so the recording is not sent to the box for transcription (it still goes to the box when you ask for speakers to be told apart).
• Your exact spoken words are kept as a quote when the agent turns them into a card or note, not smoothed into the agent's own phrasing. See provenance.md.

What it needs

A transcription vendor key for accurate, high-quality transcription: OpenAI, Mistral, or Deepgram. A microphone, in the browser or on the phone. The iPhone app for on-device speech recognition. A text-to-speech provider for spoken replies: OpenAI, or Gemini with a Google Gemini key. On desktop browsers a spoken reply starts playing as its first part arrives rather than after the whole clip is made. See ../install/index.md.

How it works, briefly

Recording happens in the browser or the iPhone app. The recording is transcribed by the vendor key you've configured, or on-device on the phone, and the transcript becomes a chat message or a card. A spoken reply comes from a delivery instruction the agent writes into its response, naming the voice and how to say it; the box turns that into audio.

Limits

Per-message voice overrides only apply in chat, not in background jobs or procedures. The documentation does not describe offline (no-network) voice support.

Go deeper

chat.md, phone-capture.md, provenance.md, ../reference/chat-voice.md, ../reference/narration-mode.md, ../reference/cards/audio.md, ../reference/cards/capture-session.md