Skip to main content
Draft: these steps await verification against the current release of AgentKit.
Outcome: you speak to an agent instead of typing: either your words appear in the composer as text (dictation), or you hold a call where a voice model talks with you and hands your words to the session (voice mode).

What you need

  • A session AgentKit runs, such as one started with New on an agent’s page.
  • For dictation: a Deepgram API key.
  • For voice mode: a Gemini Live or GPT Live key.
  • Microphone permission for AgentKit; macOS asks on first use.

Steps in the app

1

Add your Deepgram key

Open Settings > Transcription and paste your Deepgram API key. AgentKit validates it. Optionally choose the model, the language (Automatic by default), the dictation key (right Option by default), key terms and sounds.
2

Start listening

In a session’s composer, press the microphone beside Send, or tap the dictation key once. The start sound plays once the microphone is listening, and the button shows the level of your voice.
3

Speak

Text appears at the cursor as you speak, dimmed while it is still being recognised.
4

Stop or send

Press the microphone or the key again to stop, Escape to stop and keep the text, or Send to wait for your last words and send. Listening stops by itself after 30 seconds of silence.

Check it worked

  1. Dictation: your words appear in the composer and settle as ordinary text.
  2. Voice mode: the button runs a light round its outline and a glow follows the voice; after a handover, the session’s feed shows your message headed Voice mode messages with a line per utterance, and the agent answers it.

Common failures