ai-music-mcp

The only local AI music production MCP -- generate, separate stems, convert voice, full pipeline. No API costs.

---

MusicGen instrumental generation, Demucs stem separation, and RVC voice conversion -- orchestrated through a single MCP server that runs entirely on your machine. No cloud APIs, no rate limits, no per-track charges. Feed it a text prompt, get back a produced track with isolated stems and converted vocals.

  Prompt                                                     Final Output
    |                                                             |
    v                                                             v
+----------+     +----------+     +-----------+     +-----------+
| MusicGen | --> | Demucs   | --> | RVC Voice | --> | Mixed     |
| Generate |     | Separate |     | Convert   |     | Audio     |
+----------+     +----------+     +-----------+     +-----------+
  "dark trap       vocals.wav      converted-        Ready to
   beat with       drums.wav       vocals.wav         use
   808s"           bass.wav
                   other.wav

Install

npm install -g ai-music-mcp

Configure

Add to claude_desktop_config.json or ~/.mcp.json:

{
  "mcpServers": {
    "ai-music": {
      "command": "ai-music-mcp",
      "env": {
        "MUSIC_STUDIO_DIR": "/path/to/ai-music-studio",
        "RVC_DIR": "/path/to/ai-music-rvc",
        "MUSIC_OUTPUT_DIR": "~/Desktop/AI-Music"
      }
    }
  }
}

Environment Variables

Variable	Required	Default	Description
MUSIC_STUDIO_DIR	Yes	--	Path to ai-music-studio with MusicGen/Demucs scripts and Python venv
RVC_DIR	No	--	Path to RVC WebUI installation (needed for voice conversion)
MUSIC_OUTPUT_DIR	No	~/Desktop/AI-Music	Output directory for all generated files
MUSIC_GENERATE_TIMEOUT	No	600000	MusicGen timeout in ms (10 min)
MUSIC_STEMS_TIMEOUT	No	600000	Demucs timeout in ms (10 min)
MUSIC_CONVERT_TIMEOUT	No	300000	RVC timeout in ms (5 min)

Tool Reference

Tool	Description
music_generate	Generate instrumental music from a text prompt via MusicGen (small/medium/large, 1--30s)
music_separate_stems	Separate any audio into vocals, drums, bass, and other stems via Demucs
music_convert_voice	Convert vocals to a target voice with RVC -- pitch shifting, F0 method selection
music_list_voice_models	List trained RVC voice models (.pth) and their index files
music_full_pipeline	End-to-end: generate + separate stems + voice convert in one call (5--15 min)
music_list_files	List all generated audio by category (generated, stems, converted)
music_list_stems	Browse separated stems by track name and Demucs model
music_get_info	Audio file metadata -- duration, sample rate, channels, codec, bitrate (ffprobe)

Usage Examples

> Generate a 20-second dark trap beat with heavy 808s and hi-hats
  -> music_generate { prompt: "dark trap beat...", duration: 20, model: "small" }

> Separate the stems from ~/Music/song.mp3
  -> music_separate_stems { input_file: "~/Music/song.mp3", model: "htdemucs_ft" }

> Convert the vocals to matthew's voice, shift pitch up 2 semitones
  -> music_convert_voice { input_file: ".../vocals.wav", model_name: "matthew.pth", pitch: 2 }

> Full pipeline: lo-fi chill beat, separate it, apply my voice model
  -> music_full_pipeline { prompt: "lo-fi chill beat with warm piano", voice_model: "matthew.pth" }

Output Structure

~/Desktop/AI-Music/
├── dark-trap-beat-808s.wav           # Generated instrumentals
├── lofi-chill-beat-piano.wav
├── stems/
│   └── htdemucs/
│       └── dark-trap-beat-808s/
│           ├── vocals.wav            # Isolated vocals
│           ├── drums.wav             # Isolated drums
│           ├── bass.wav              # Isolated bass
│           └── other.wav             # Everything else
└── converted/
    └── converted-vocals-dark-trap.wav  # RVC voice conversion output

Prerequisites

Dependency	Purpose	Install
Node.js 18+	MCP server runtime	nodejs.org
Python 3.11+	ML model execution	brew install python@3.11
PyTorch	Neural network backend	pytorch.org
ffmpeg / ffprobe	Audio processing and metadata	brew install ffmpeg
ai-music-studio	MusicGen + Demucs Python scripts	Clone and set up venv
RVC WebUI	Voice conversion (optional)	Clone and set up venv

Why This One?

	ai-music-mcp	musicmcp.ai	MiniMax Music API
Runs locally	Yes -- zero network calls	No -- cloud only	No -- cloud only
Per-track cost	$0	Paid per generation	Paid per API call
Stem separation	Built-in (Demucs)	No	No
Voice conversion	Built-in (RVC)	No	No
Full pipeline	Generate + stems + voice in one call	Generate only	Generate only
Model control	Choose MusicGen small/medium/large, Demucs model, RVC model	Fixed	Fixed

The only MCP server that chains generation, stem separation, and voice conversion into a single local pipeline. Every other option is cloud-only and charges per track.

Development

cd typescript
npm install
npm run build        # Compile + bundle
npm run watch        # TypeScript watch mode
npm run inspector    # Launch MCP Inspector for interactive testing

Related Projects

• ai-music-studio -- Python CLI for MusicGen and Demucs (backend)
• rvc-mcp -- MCP server for RVC model training and management
• music-distro -- MCP server for distributing tracks to SoundCloud and YouTube

License

MIT -- Purple Squirrel Media