muse-tts-embed

What Is This?

MUSE TTS Embed gives Claude a persistent audio player — right in the chat. Everything runs on your machine.

Looking for direct audio playback without a player widget? See MUSE TTS Live.

Related MCP server: MCP Audio Server

Features

• Embedded Player — play, pause, seek, replay

• Voice Selector — switch between 54 voices without leaving the conversation

• Voice Cloning — drop a WAV, get a clone (~7s)

• Speed Control — 0.5x through 2.0x playback

• Download — save any generation as WAV or MP3

• 3 Engines — Kokoro-82M (~1s), IndexTTS-1.5 (Apple Silicon cloning), Chatterbox (cross-platform cloning)

• Fully Local — no cloud APIs, no Docker, no subscription

Quick Start

1. Install dependencies

macOS (Apple Silicon — fastest):

pip install fastmcp mlx_audio

Windows / Linux / Intel Mac:

pip install fastmcp kokoro soundfile numpy

On Linux, you also need espeak-ng: sudo apt install espeak-ng

2. Add to Claude Desktop

Open Settings > Developer > Edit Config and add:

{
  "mcpServers": {
    "muse-tts-embed": {
      "command": "python3",
      "args": ["/path/to/muse-tts-embed/server.py"]
    }
  }
}

Restart Claude Desktop.

3. Speak

Ask Claude to speak anything. Try: "Say hello in a warm voice" or "Read this paragraph aloud"

Voice Cloning

Add your own reference WAV files to the voices/ directory. They'll be detected on startup.

voices/
  my_narrator.wav
  interview_voice.wav

Then ask Claude: "Speak this using the my_narrator clone"

Or use ref_audio to point to any WAV in voices/ or ~/Downloads/:

"Read this aloud using the reference audio at ~/Downloads/sample.wav"

Configuration

HTTP Mode (Web / Mobile)

For Claude Web or Mobile, run the server in HTTP mode behind a tunnel:

export MUSE_AUTH_TOKEN=your-secret-token
python3 server.py --http

Then expose via tunnel (ngrok, cloudflared) and add the URL to Claude's MCP settings.

Browsers block auto-play. The player shows a "click to play" hint — tap play to start.

Voices

54 preset voices across 9 languages (American English, British English, Spanish, French, Hindi, Italian, Japanese, Portuguese, Mandarin).

Tools

How It Works

Audio is delivered via structuredContent — bypasses model context, no size limit. The player is a self-contained HTML/JS app.

Requirements

• Python 3.10+

• Claude Desktop (latest version with MCP Apps support)

• One of: mlx_audio (Mac M-series) or kokoro + soundfile (any platform)

• ~200MB disk space (model downloads on first use)

Player shows "Generating..." but nothing happens: Check that the TTS engine is installed. Run muse_embed_check to verify status.

No sound on Web/Mobile: Browsers block auto-play. Click the play button.

"No TTS engine found": See installation above.

Model download is slow: First run downloads ~200MB.

"Text too long" error: Max 2000 characters per generation (~2 minutes of speech). Break longer text into parts.

Player doesn't appear: Update Claude Desktop to the latest version with MCP Apps support.

License

Licensed under the Apache License, Version 2.0.

Copyright 2026 The Funkatorium (Falco & Rook Schäfer). Protected under German Copyright Law (Urheberrechtsgesetz). Jurisdiction: Amtsgericht Berlin.