Docs

Command Palette

Search for a command to run...

Docs/AI Conversations/Voice Mode

Voice Mode

Talk to Wyatt in real time with a persistent voice agent that follows you across pages.

Overview

Voice Mode lets you talk to Wyatt naturally while keeping the same conversation context as text chat. You can start voice from the header, the conversation page, or the conversations home screen.

How It Works

  1. Start voice from any supported entry point.
  2. A voice island appears at the top of the app and stays with you across routes.
  3. Speak naturally. The agent can discover and invoke the same authoritative Wyatt workspace actions used by chat, while keeping the live voice tool surface compact.
  4. Stop anytime from the island.

Controls

  • Mute: Toggle your microphone on/off.
  • Stop: End the active voice session.
  • Open conversation: Jump to the linked conversation thread.

Conversation History

Voice interactions are saved as normal messages in your conversation history. Spoken turns are displayed in italics with lower contrast so they are visually distinct from typed turns.

  • Starting voice from an existing conversation reuses that same conversation thread.
  • Starting voice from the header, home screen, or other global entry points creates a new conversation with context_type="voice".
  • Voice turns are stored as standard conversation messages with voice metadata so they stay synced with the same chat history as typed messages.

Settings

In Settings → AI, you can:

  • Enable/disable voice mode
  • Choose a default ElevenLabs voice
  • Voice choice is applied when you start a session (if you change it mid-session, restart voice mode to apply)

Background Behavior (Web)

Voice Mode is best-effort while running in a browser:

  • Works best when a Wyatt tab stays active.
  • Hidden tabs may be throttled depending on browser/OS.
  • Mobile browser background and screen-off behavior is limited.
  • For local development, the custom LLM callback URL must be publicly reachable (for example via a tunnel).
  • The callback URL must be Wyatt’s /api/voice/elevenlabs/custom-llm endpoint so tool and connector access is preserved.

Supported Runtime

The repaired web voice flow is supported on the ElevenLabs server-model path. That is the primary runtime for persistent voice conversations stored in chat history.

Voice sessions now register the compact discovery/invoke tool surface with the active workspace conversation, rather than exposing a separate voice-only capability universe.

If you need always-on screen-off voice, use a native app experience when available.