Voice Commands

The agent chat input has a microphone button. Speak instead of typing.

How to use it

1. Click the microphone icon next to the Send button.
2. Grant microphone permission the first time (macOS or Windows system prompt).
3. Speak. A live waveform shows recording.
4. Click the microphone again, or pause for a moment, to stop.
5. The audio is transcribed and the text appears in the chat input. Send normally.

The voice flow is the same as the generation panels’ voice input — it uses Whisper or Gemini, whichever you have a key for.

What voice is good for

• Hands-free iteration — describe edits while watching the canvas.
• Long descriptive prompts — speaking is faster than typing once the prompt gets past two sentences.
• Brainstorming aloud — let the transcription capture the thought, then edit before sending.

What voice is not great for

• Specific element names — transcribers stumble on stylised file names. Use @-mention after transcription.
• Code or markup — the LLM-friendly direction is to describe results, not paste code.

Privacy

• Audio is sent only to the configured transcription provider (OpenAI or Google).
• No audio is stored on disk after transcription.
• Microphone permission can be revoked at any time in OS privacy settings.

Combine voice and mention

A common pattern:

1. Click microphone, say: “Generate a wide-shot landing hero with the product visible on the right and a soft gradient background.”
2. Stop recording. The transcript fills the input.
3. Type @ and pick the product image from the popover.
4. Send.

Where to go next

• Giving Commands
• Mentioning Canvas Media
• Voice Input — same feature in generation panels.