Once voice is configured, tap the microphone icon in any session to start a voice conversation.

​
How the voice agent thinks

The voice agent uses a priority-based decision framework:
  1. Tool action? — If you’re asking to do something in the app (switch sessions, approve permissions), it calls a tool directly
  2. For your coding agent? — If you’re giving coding instructions, it forwards to Claude/Codex/Gemini
  3. For the voice agent itself? — If you’re chatting or asking a question, it replies directly
  4. Uncertain? — It asks a clarifying question rather than guessing

​
Example commands

Forwarded to your coding agent:
  • “Refactor the authentication module to use JWT”
  • “Write tests for the user registration endpoint”
  • “Explain what the sync function does”
Handled directly by voice agent:
  • “What’s the current session status?”
  • “Read me the latest agent response”
Tool actions:
  • “Switch to session 2”
  • “Approve the permission request”
  • “Create a new Claude session”
  • “Go back to the home screen”
  • “End this voice conversation”

​
Speech correction

The voice agent automatically corrects common speech-to-text errors, especially for technical terms. For example:
  • Homophones: “组建” → “组件” (component in Chinese)
  • Technical terms: Code function names and programming concepts are corrected using context from your current session

​
Context awareness

The voice agent knows about:
  • Your current session and its recent messages
  • What the AI coding agent is currently doing
  • Which files are being edited
This context helps it make better decisions about how to handle your requests. For example, if your agent just finished a task, the voice agent can summarize the results without you having to ask.

​
Tips

  • Keep it short — The voice agent responds best to concise instructions
  • Be specific — “Fix the bug in auth.ts” is better than “fix that thing”
  • Use natural language — You don’t need to use special syntax or commands