Voice Mode

# Voicemode Resource Structure Plan ## Goal Reduce the `mcp__voicemode__converse` tool description from ~4000 tokens to ~1000 tokens by moving detailed documentation into MCP resources that can be fetched on-demand. ## Current State - Tool description: ~4000 tokens - After initial reduction: ~1000 tokens - Includes extensive examples, patterns, and parameter documentation ## Target State - Tool description: ~200-300 tokens (minimal) - Detailed docs: Separate MCP resources (~50-100 tokens per resource in listing) - LLM fetches resources only when needed ## Proposed Tool Description The minimal tool description should include: ``` Have an ongoing voice conversation - speak a message and optionally listen for response. See MCP resources for detailed documentation: - voicemode-quickstart: Basic usage examples - voicemode-parameters: Detailed parameter explanations - voicemode-languages: Non-English language support - voicemode-patterns: Best practices and conversation patterns - voicemode-troubleshooting: Audio, VAD, and connectivity issues Key parameters: - message (required): The message to speak - wait_for_response: Listen for response after speaking (default: true) - listen_duration: Max listen time in seconds (default: 120) - voice: TTS voice name (auto-selected unless specified) - tts_provider: openai or kokoro (auto-selected unless specified) - disable_silence_detection: Disable auto-stop on silence (default: false) For full parameter list and advanced options, see voicemode-parameters resource. ``` ## Resource Structure This directory mirrors the proposed MCP resource structure: - `quickstart.md` - Basic usage and common patterns - `parameters.md` - Complete parameter reference with descriptions - `languages.md` - Non-English language support guide - `patterns.md` - Advanced patterns (parallel operations, etc.) - `troubleshooting.md` - Common issues and solutions ## Benefits 1. **Token Efficiency**: Only load documentation when needed 2. **Better Organization**: Separate concerns into logical chunks 3. **Easier Maintenance**: Update docs without changing tool definition 4. **Scalability**: Can add more resources without bloating tool context 5. **Smart Discovery**: Resource names in listing help LLM find what it needs ## Implementation Notes - Resource names should be self-explanatory - Resource URIs should follow pattern: `voicemode://docs/{resource-name}` - Tool description should hint at resource availability - LLM will see resource list in context and fetch as needed