voice-start

name	voice-start
description	Start a voice conversation with Claude about your codebase. Talk through your mic, get answers through your speakers. Free, no telephony needed.
user-invocable	true
disable-model-invocation	true
allowed-tools	["mcp__codebase-voice__start_local_session","mcp__codebase-voice__poll_question","mcp__codebase-voice__provide_answer","mcp__codebase-voice__stop_session","mcp__codebase-voice__get_config","Read","Grep","Glob"]

Voice Start — Talk to Claude About Your Code

Start a local voice session. Ask questions about your codebase through your microphone and hear answers through your speakers. No telephony setup needed.

Requirements

• Deepgram API key (free tier: 12,000 min/month)
• A microphone

Steps

1. Check configuration: Call mcp__codebase-voice__get_config to verify Deepgram API key is set. If not configured, tell the user to run /voice-config first and stop.

2. Index the codebase: Quickly scan the current project to build context:

   • Use Glob with **/*.{py,js,ts,go,rs,java,rb,cpp,c,h} to get source files (limit to first 50)
   • Use Read to read: README.md, CLAUDE.md, package.json or pyproject.toml (whichever exists)
   • Build a 2-3 sentence codebase summary including: language, framework, purpose, key directories

3. Start the session: Call mcp__codebase-voice__start_local_session with the project path and codebase summary.

4. Enter the question-answer loop: Continuously poll for voice questions and answer them:

   LOOP:
     a. Call mcp__codebase-voice__poll_question (timeout_ms: 2000)
     b. If result is "__NO_QUESTION__" → go to step a (keep polling)
     c. If result is "__LEAVE__" → call stop_session and exit
     d. If result is "__SESSION_ENDED__" → exit
     e. Otherwise, the result is a question about the codebase:
        - Use Grep and Read to search the codebase for relevant code
        - Formulate a concise answer (1-3 sentences, conversational tone)
        - Call mcp__codebase-voice__provide_answer with the answer
        - Go to step a
   

   IMPORTANT: Keep answers concise and conversational. The user is listening, not reading. Say things like "The auth module is in src/auth.py, it uses JWT tokens for session management" not long code blocks or technical documentation.

5. On exit: Call mcp__codebase-voice__stop_session and tell the user the session ended.