The live assistant

The live assistant lets an AI take part in a meeting by voice. An agent in your own Claude Code session follows the transcript of the recording in progress, decides when it has something worth adding, and speaks a short reply through a local voice on this Mac. The model is whichever one you run in Claude Code, on your own subscription; Ambient needs no API key for it.

It is off until you turn it on, and when it starts watching a meeting it first says out loud that an AI is listening and may speak.

It is early software. The transcript goes to the model you run in Claude Code; read what leaves this Mac before using it with anyone who has not agreed to that.

Setting it up

Connect Ambient to Claude Code once, as in MCP setup:

claude mcp add ambient -- "$(which ambient)" mcp

Turn the assistant on, from the status menu (Live Assistant, which shows a tick when it is on) or with:
```
ambient config assistant on
```

Using it in a meeting

Start recording, then in Claude Code pick a fast model and run the watch prompt:

claude --model haiku
> /mcp__ambient__watch

Anything you type after the command is passed on, for example /mcp__ambient__watch Listen out for budget numbers and correct any that are wrong.

The agent then:

Starts watching. Ambient says out loud: “Just so you know, an AI assistant called Claude is listening to this meeting and may speak up.” While it watches, the status menu shows AI assistant listening — it may speak.
Waits for each burst of speech, reads it, and decides. It speaks when someone addresses it by name, when a factual question put to the room goes unanswered, or when something clearly wrong is about to be acted on. Otherwise it stays quiet.
Speaks one to three short sentences through the local voice.
Stops when the recording stops, when the assistant is turned off, or when you tell it to.

Ambient, not the agent, enforces a cooldown after each utterance and a cap per meeting, so a model that loses the thread cannot talk over a meeting. Its own voice, picked up again by the microphone, is left out of what the agent reads. Turning Live Assistant off in the menu silences it immediately, even mid-meeting.

Haiku is fast enough to answer about three seconds after its name is said. A larger model answers more slowly but may judge better when to speak.

The voice

Speech runs in a separate helper process, so a voice model never shares memory with the recorder. The helper lives in voice/ in the source tree and runs through uv; its first start installs the chosen engine’s runtime and downloads its model.

voice.engine is set per Mac, and auto picks by installed memory:

Engine	Picked by `auto`	First sound	Helper peak memory	Expression
`qwen3-tts` (Qwen3-TTS 0.6B, MLX)	32 GB or more	~0.05–0.1 s (streamed)	~2.7 GB preset, ~1.6 GB cloned voice	A delivery instruction per reply (preset), or a designed voice
`supertonic3` (Supertonic 3, ONNX)	16–31 GB	~0.4 s (first sentence)	~0.6 GB	`<laugh>` / `<sigh>` on amused or apologetic replies
`kokoro` (Kokoro-82M int8, ONNX)	under 16 GB	~0.7 s (first sentence)	~0.4 GB	None

The figures were measured on an M5 Max; see the design note. The agent picks a tone for each utterance (neutral, warm, amused, excited, apologetic or concerned). With a preset Qwen3-TTS speaker it becomes tone of voice, and voice.style adds a standing instruction to every reply, for example ambient config voice.style "Dry and understated.".

A voice of your own

Qwen3-TTS can also speak in a voice designed from a description. The voice is designed once by the larger 1.7B VoiceDesign model, then a small clone of it speaks every reply:

ambient config voice.reference voices/v1-gravel

That selects the voice shipped in voice/voices/v1-gravel: an older Scottish man with a deep, gravelly voice. Its voice.json records the description and delivery cue it was designed from. To design another, set a description and, optionally, a cue, and leave voice.reference unset:

ambient config voice.description "A warm, softly spoken woman in her forties with a slight Irish lilt."
ambient config voice.cue "She delivers this calmly, with a smile."

The first start then designs the voice, which downloads the 1.7B model and briefly uses about 5.6 GB. Later starts reuse the design. A cloned voice sounds the same on every reply. It does not take a per-reply mood, so the reply’s emotion comes through only in its words.

A cloned voice is lighter than a preset one. The helper keeps a slimmed copy of the model, built on first use. It peaks at about 1.6 GB, settles at about 1.1 GB, and speaks within about 50 ms.

Licences differ: Qwen3-TTS and Kokoro are Apache-2.0; Supertonic 3 is OpenRAIL-M, which carries use restrictions. Models download to the Hugging Face cache, ~/.cache/supertonic3 and ~/Library/Caches/Ambient/voice, not into the app. Designed voices and the slimmed clone are cached in ~/Library/Caches/Ambient/voice too, about 0.9 GB per voice.

Settings

All are set with ambient config <key> <value>; see settings.

Key	Default	Effect
`assistant`	`off`	The on/off switch.
`assistant.name`	`Claude`	What people call it, and what it answers to.
`assistant.cooldown_s`	`60`	Seconds of silence after each utterance.
`assistant.max_per_meeting`	`10`	Most utterances in one meeting.
`voice.engine`	`auto`	`auto`, `qwen3-tts`, `supertonic3` or `kokoro`.
`voice.speaker`	engine default	`Vivian` (Qwen3-TTS), `F1` (Supertonic), `af_heart` (Kokoro).
`voice.style`	none	Standing delivery instruction for preset Qwen3-TTS speakers.
`voice.reference`	none	A designed voice folder to clone, such as `voices/v1-gravel`. Relative paths are inside the helper folder. Qwen3-TTS only.
`voice.description`	none	Design a voice from this description, once. Used when `voice.reference` is unset.
`voice.cue`	none	A delivery cue appended to the description when designing.
`voice.helper_dir`	`voice/` in the checkout the binary runs from	Where the helper project is.

To hear the voice without playing it, for a test, start the server as ambient mcp --no-play --save-audio <dir>; each utterance is written to <dir>/reply-<n>.wav.

What leaves this Mac

Audio never does. While the agent watches, each new stretch of transcript (who spoke, when, and what they said) goes to the agent, and from there to the model it runs. With Claude Code that is Anthropic, under the terms of your Claude plan. Ambient itself sends nothing anywhere: it makes no model call and holds no key.

Each utterance is appended to assistant.jsonl in the session folder, so what the assistant said is kept with the meeting and deleted with it.

Cost

The agent runs on your Claude Code plan, so a meeting uses your plan’s usage, not a separate bill. Ambient groups speech into bursts, so the agent takes one short turn per burst rather than one per line. A 75-second sample meeting took 13 turns with Haiku. Claude Code reported that as $0.006 in API terms, which is an equivalent figure, not a charge on a subscription.