MusicMixCode — Ableton Auto-Mix MCP

🇷🇺 Читать на русском

An MCP server for auto-mixing and auto-mastering in Ableton Live tuned to a musical style/genre. An AI agent (opencode, Claude Code, Claude Desktop...) analyzes render files of your tracks, compares them against a style profile and produces/applies corrections: levels, pan, EQ hints, compression — and additionally renders a mastered preview mix (sidechain, HPF, mud-cut, soft-clipper, true-peak limiter).

Features

• 10 MCP tools: analysis, style-based auto-mix, preview render, release check.
• 11 style profiles (JSON): techno, hip_hop, pop, lo_fi, ambient, balanced, trance, breaks, dubstep, drum_n_bass, trap.
• Mastering stage: sidechain (kick → bass, snare-band Dynamic EQ), role-based HPF, 200–500 Hz mud-cut, tanh soft-clipper, LUFS normalization, true-peak lookahead limiter (4× oversampling), TPDF dither.
• Stereo imaging: mid/side width per role (mono for kick/sub, wide/very_wide for hats/pads) + panning.
• Spectral role detection: if the file name doesn't hint a role, it's inferred from the spectrum.
• Release Check: LUFS/LRA/true-peak/RMS/sub-mid-gap against top-label targets — ready / needs_work verdict.
• Conflict analysis: which track pairs fight over frequency bands.
• Offline mode: analysis and previews work without Ableton Live (only WAV renders are needed).

Architecture

You → MCP client → ableton-auto-mix-mcp (MCP server)
                         ├── styles/*.json     — style profiles (target curves)
                         ├── analyzer.py       — LUFS/LRA, spectrum, stereo width (librosa/pyloudnorm)
                         ├── mixer.py          — engine: analysis vs profile → corrections (anchor = kick, LUFS)
                         ├── preview.py        — preview-mix render + mastering chain
                         ├── qa.py             — conflict analysis + release check
                         └── ableton_client.py — AbletonOSC (python-osc) → Ableton Live

Principle: the MCP itself does not "mix" — it provides tools and metrics while the model decides. Cycle: render → analyze → dry-run report → preview mix → release check → apply corrections.

Installation

pip install -r requirements.txt        # or: pip install -e .

Then connect the MCP server in your client (example for Claude Code / opencode):

{ "mcpServers": { "ableton-auto-mix": {
    "command": "python", "args": ["-m", "ableton_auto_mix"],
    "cwd": "C:/path/to/ableton-auto-mix-mcp"
}}}

Analysis and preview rendering do not require Ableton Live — one WAV per track is enough.

Ableton setup (optional, for auto-apply)

1. Launch Ableton Live.
2. Install the AbletonOSC control surface (https://github.com/ideoforms/AbletonOSC).
3. Preferences → Link, Tempo & MIDI → Control Surface → AbletonOSC.
4. Bounce each track into renders/ (one WAV per track) for analysis.

MCP tools

Example session

"Mix the renders from renders/ in a techno style, show what to change"
→ auto_mix("techno", "renders", true)

"OK, go ahead"
→ auto_mix("techno", "renders", false)

"Render a mastered preview mix"
→ preview_mix("breaks", "renders", max_duration=30)

"Check if the mix is release-ready"
→ release_check("breaks", "renders")

CLI (without an MCP client)

Everything is available from the command line via python -m ableton_auto_mix <command> (or ableton-auto-mix-mcp <command> after installing):

ableton-auto-mix-mcp styles                                  # list styles
ableton-auto-mix-mcp style breaks                            # style profile
ableton-auto-mix-mcp analyze renders/                        # metrics of all renders
ableton-auto-mix-mcp suggest renders/                        # which style fits
ableton-auto-mix-mcp mix breaks renders/                     # dry-run: what to change
ableton-auto-mix-mcp preview breaks renders/ --max-duration 30   # preview mix to WAV
ableton-auto-mix-mcp conflicts renders/                      # frequency conflicts
ableton-auto-mix-mcp release breaks renders/                 # ready/needs_work verdict

Output is JSON (script-friendly). Example: preview breaks renders/ --manual-gain "bass=2.0,snt2=-4.0" --output out.wav.

Styles

Styles ship inside the package (ableton_auto_mix/styles/) — 11 profiles: techno, hip_hop, pop, lo_fi, ambient, balanced, trance, breaks, drum_n_bass, trap. Each profile defines: target LUFS/LRA, a spectral curve (6 bands), relative instrument levels (kick/bass/vocals/lead/wobble/breaks/...), per-role HPF, mud-cut, sidechain, mastering settings, compression and FX recommendations. You can add your own: copy a JSON and change name and targets.

To use custom styles without editing the package, point to an env var:

export ABLETON_AUTO_MIX_STYLES_DIR=/path/to/my-styles

Tests

python tests/test_smoke.py   # 5 smoke tests: render → mix → preview → release check

Limitations

• Analysis is done on renders (offline), since Live doesn't stream samples in real time via OSC.
• A track role is inferred from its filename; unknown names fall back to spectrum analysis (kick, bass, vocals...). Name tracks explicitly for accuracy.
• auto_mix(dry_run=false) requires a running Ableton Live with the AbletonOSC control surface.
• Preview mastering applies the profile's standard chain; do the final touch-up manually.

License

MIT © 2026 MusicMixCode / HighVoltSound