msp-music-pipeline

AI Music Production Agent pipeline. Drop in a raw track, get back a mixed and mastered stereo file ready for streaming platforms.

Raw Track → Stem Separation → Analysis → Processing → Mixdown → Mastering → Done

• Separates audio into stems (drums, bass, vocals, other) using Meta's Demucs
• Analyzes each stem: BPM, key, loudness, spectral features, clipping detection
• Processes each stem with role-appropriate FX chains (EQ, compression, reverb)
• Mixes stems back together with level balancing
• Masters to -14 LUFS / -1 dBTP (Spotify, Apple Music, YouTube Music standard)

All tools are composable and stateless — use the full pipeline or individual tools.

• Python 3.11+
• ffmpeg (apt install ffmpeg / brew install ffmpeg / choco install ffmpeg)
• ~2GB disk space for Demucs models (downloaded on first run)

pip install -e .
# With AI review support:
pip install -e ".[review]"

# Full pipeline
python -m msp.runner my_track.wav ./output

# Analyze a stem only
python -m msp.runner --analyze-only stem.wav

# Process a single stem
python -m msp.runner --process-only stem.wav --role drums --out drums_processed.wav

Output (flat directory, AI-friendly):

output/
├── drums_raw.wav, bass_raw.wav, vocals_raw.wav, other_raw.wav
├── drums_processed.wav, bass_processed.wav, vocals_processed.wav, other_processed.wav
├── mix.wav, master.wav
├── analysis.json, decisions.json, pipeline_log.json
└── review.json (optional, with --review flag)

Add to your MCP config:

{
  "mcpServers": {
    "msp": {
      "command": "python",
      "args": ["-m", "msp.server"]
    }
  }
}

Then ask Claude:

• "Process track.wav through the full pipeline"
• "Just analyze the drums stem"
• "Master the mix targeting -11 LUFS for YouTube"

Every tool is **stateless and comp