kyutai-tts-mcp 0.6.0

MCP server wrapping Kyutai Pocket TTS for Claude Code (in-process, streaming, gap-free playback, multi-language)

kyutai-tts-mcp

Local-only voice for any MCP client (Claude Code, Claude Desktop, Cursor, etc.) via Kyutai Pocket TTS. No cloud, no API keys, no rate limits.

• 🇫🇷 French (Estelle), 🇬🇧 English (Alba), plus Spanish, German, Italian, Portuguese
• TTFA ~80–200 ms thanks to native streaming via the pocket-tts Python API
• Multi-language at runtime — pass language= per speak() call, models load lazily and cache
• ~4-5× real-time generation on Apple Silicon / Intel CPU
• Non-blocking speak(), gap-free playback via sounddevice write-mode
• ~600 MB venv, ~1 GB model cache per language

Install

uvx kyutai-tts-mcp --help

Or persistent:

uv tool install kyutai-tts-mcp

Then add to your MCP client's .mcp.json:

{
  "mcpServers": {
    "kyutai-tts": {
      "command": "uvx",
      "args": ["kyutai-tts-mcp", "--language", "french_24l"]
    }
  }
}

Replace french_24l with english, spanish_24l, german_24l, italian_24l, or portuguese_24l for your default language. Per-call language= overrides this default.

MCP tools

Tool	Purpose
speak(text, voice?, language?, interrupt?)	Generate audio for text and queue it for background playback. Returns immediately, streaming generation. By default, calls queue and play sequentially (including across turns). Pass interrupt=True to abort current playback and clear the queue first. Pass language= to switch model on the fly.
stop_speaking()	Stop current playback, drop queue, cancel in-flight generation. Use for explicit "mute" requests. For mid-turn interruption + new speech, use speak(..., interrupt=True) instead.
status()	Report loaded languages, queue depths, sample rate, last error.

Configuration

The default language is set via --language (CLI) or KYUTAI_TTS_LANGUAGE (env var). Other knobs (all env):

• KYUTAI_TTS_VOICE — default voice name
• KYUTAI_TTS_DEVICE — PyTorch device (default cpu; mps not supported)
• KYUTAI_TTS_QUANTIZE — set to 1 for int8 quantization
• KYUTAI_TTS_MAX_TOKENS — max tokens per streaming chunk (default 50)

Voice cloning

Pocket-tts can clone a voice from a short audio sample. The cloning runs 100 % locally — your audio never leaves the machine — but the cloning-enabled model checkpoint is gated on Hugging Face. One-time setup:

1. Accept the terms at huggingface.co/kyutai/pocket-tts
2. uv tool install huggingface-hub && hf auth login
3. rm -rf ~/.cache/huggingface/hub/models--kyutai--pocket-tts to force re-downloading the cloning-enabled checkpoint

Then either pass the audio file directly to speak(voice="/path/to.wav"), or pre-extract it as a .safetensors for instant loading:

uvx kyutai-tts-mcp extract-voice \
  --audio ~/voices/yours.wav \
  --out ~/voices/yours.safetensors \
  --language french_24l \
  --truncate

Then speak(voice="/path/to/yours.safetensors"). Full guide and recording recommendations: main repo.

Claude Code users

If you're on Claude Code (CLI, desktop, or via Cursor), you can install the bundled plugin (MCP wiring + /voice-mode skill) in one shot:

/plugin marketplace add Vincweb/kyutai-tts-mcp
/plugin install kyutai-tts@vincweb-tools

See the main repo for architecture details, voice catalog, and a side-by-side comparison with voxtral-mcp.

Requirements

• macOS (Apple Silicon recommended; Intel works — the model is CPU-only)
• Python 3.10 – 3.13
• ~1 GB free RAM per loaded language

License

MIT. Kyutai pocket-tts itself is under its own permissive licence (see upstream).