Talk to Claude Code while your hands are full

What it does

VoiceMode is an MCP server that lets you hold spoken conversations with Claude Code (or any MCP-capable agent). You talk, it transcribes, the agent responds, and you hear the answer back — with silence detection so it stops listening when you stop talking. It’s aimed squarely at the moments when typing isn’t an option: walking between meetings, debugging over a stove, or resting screen-tired eyes.

The interesting bit

The local speech stack — Whisper.cpp for speech-to-text and Kokoro for text-to-speech — exposes the same API as OpenAI’s, so the project swaps between cloud and local backends without changing anything else. That’s a clean bit of glue: privacy mode is a backend choice, not a fork.

Key highlights

• Runs fully offline with local Whisper STT and Kokoro TTS
• Works as a Claude Code plugin or a standalone MCP server
• Smart silence detection ends recording when you stop speaking
• Broad platform support: Linux, macOS, Windows (WSL), NixOS; Python 3.10–3.14
• Debug-friendly: VOICEMODE_SAVE_AUDIO archives every conversation to disk

Caveats

• Local voice services need platform-specific system dependencies (portaudio, ffmpeg, and friends), and WSL2 users need pulseaudio packages for microphone access — the “works out of the box” claim has an asterisk on some setups.
• Cloud mode requires an OpenAI API key as fallback.

Verdict

Worth a look if you use Claude Code and want hands-free sessions or a fully local voice pipeline. If you don’t live in an MCP-capable agent, there’s not much here for you.