Chatterbox TTS MCP Server

A FastMCP server that exposes Chatterbox TTS capabilities over the network. Connect from Claude Code on your laptop to generate speech on your GPU-equipped desktop.

Features

• Text-to-Speech: High-quality speech synthesis with multiple models
• Voice Cloning: Clone any voice with 5-15 seconds of reference audio
• Voice Management: Save, list, and reuse voice references by name
• Long Text Support: Auto-chunking for texts longer than 40 seconds
• Remote Access: Generate speech from any machine on your network
• Download URLs: Fetch generated audio via HTTP

Models

Installation

# Clone the repo
git clone https://github.com/yourusername/chatterbox-mcp.git
cd chatterbox-mcp

# Install dependencies
pip install -r requirements.txt

# Install wsproto for websocket support (no deprecation warnings)
pip install wsproto

Configuration

Edit server.py to set your paths:

OUTPUT_DIR = Path(r"C:\path\to\output")
VOICES_DIR = Path(r"C:\path\to\voices")

Usage

Start the server

python server.py
# or
run_server.bat

Connect from Claude Code

Add to your MCP config (~/.claude/settings.json or .claude/settings.json):

{
  "mcpServers": {
    "chatterbox": {
      "type": "url",
      "url": "http://<your-pc-ip>:8765/mcp"
    }
  }
}

Tools

text_to_speech

Generate speech from text.

# Basic TTS
text_to_speech(text="Hello, world!")

# With voice cloning
text_to_speech(text="Hello", voice_name="david")

# Turbo with paralinguistic tags
text_to_speech(text="That's hilarious! [laugh]", model="turbo", voice_name="david")

# Expressive/dramatic
text_to_speech(text="...", exaggeration=0.7, cfg_weight=0.3)

list_voices

List all saved voice references.

delete_voice

Delete a saved voice reference.

Tips

• Default settings (exaggeration=0.5, cfg_weight=0.5) work well for most prompts
• Fast speakers: Lower cfg_weight to ~0.3 for better pacing
• Expressive speech: Use exaggeration=0.7, cfg_weight=0.3
• Voice cloning: Use 5-15 seconds of clean speech, minimal background noise
• Long text: Automatically split into chunks - no manual splitting needed

Requirements

• Python 3.10+
• CUDA-capable GPU (recommended)
• ~4GB VRAM for standard model

License

MIT