Ableton Live MCP
Give an AI assistant the whole of Ableton Live, not just a handful of commands.
An MCP server and Live Remote Script that expose the complete Live Object Model, plus music-theory generators and an audio "ear" that measures what the set actually sounds like.
Built on ahujasid/ableton-mcp (MIT), whose history is kept in this repository. This fork adds a generic bridge to the Live Object Model, about 40 new tools, workflow prompts, and fixes found while testing against Live 12.3.
What you can ask for
- "Give me an overview of the set, then put a i–VI–III–VII progression in A minor on the pad track with smooth voicings."
- "Write a trap beat on the drum rack, swing it a little, and humanize the hats."
- "Pull the bass down 3 dB, send 20% of it to reverb, and sidechain-style duck the pad with a volume envelope."
- "Set the filter cutoff on the Operator to 800 Hz and draw a 2-bar sweep up to 5 kHz."
- "Record bars 9–16 of the master and tell me how loud it is and whether the low end is muddy."
- "Watch what I change in Live for the next minute and tell me what I did."
- "What chords are in the Keys clip? Write a walking bass that follows them."
- "Play the drop and tell me which tracks are too hot."
- "Put a vocal chain on the Vox track and duck the bass from the kick."
- "Build me an 808 kit from my own samples and write a trap beat on it."
- "Aksak usulünde bir darbuka ritmi ve Hicaz makamında bir melodi yaz."
- "Save a snapshot, make the mix brighter, then let me A/B the two."
- "Which parts are fighting in my drop? Compare it with this reference track."
Features
76 tools in the default profile (104 with ABLETON_MCP_PROFILE=full), plus
workflow prompts: mix_check, beat_from_scratch, follow_the_harmony,
arrange_from_loop, release_check, explain_my_set, match_reference.
Audio analysis
render_and_analyze and analyze_audio_file report integrated loudness
(ITU-R BS.1770 LUFS, verified against the reference tone), true peak, loudness
range, crest factor, clipping, DC offset, energy per band (sub → air), spectral
centroid, stereo correlation and width (including below 150 Hz), and tempo and
key estimates — plus plain-language observations such as "Heavy low-mids
(250–500 Hz); a source of muddiness".
Tool profiles and safety
Every tool carries MCP annotations (read-only, destructive, idempotent), so clients can run reads without asking and confirm deletions. Failures come back as MCP tool errors with the reason, so the assistant can correct its next call. Tool calls run off the event loop, so a long Live operation does not stall the server, and the recording tools report progress. Built on the MCP Python SDK 2.x; clients on the 2025-11-25 and 2026-07-28 protocols both work.
The original upstream tools that newer ones replace (set_tempo,
fire_clip, get_clip_notes, create_midi_track, …) are left out by
default to keep the tool list short. Set ABLETON_MCP_PROFILE=full to get
them back.
Built with it
reference-dissector takes a reference track apart: the tempo grid, the tuning, every drum hit, the exact samples from your own library and a synthesized 808. It rebuilds the track in Live through this server's tools, with the original stems muted underneath for A/B. Its docs also cover the Live behaviour measured along the way and how a model that cannot hear can still tell whether two sounds are the same.
Install
Requirements: Ableton Live 11 or 12 (the Trial works),
uv (brew install uv).
1. Add the server to your client
Download ableton-live-mcp-<version>.mcpb from the
Releases page and
open it, or drag it into Settings → Extensions. Claude Desktop installs the
Python dependencies with uv the first time it starts the server. To build the
file yourself: uv run python packaging/build_mcpb.py (needs Node.js).
/plugin install ableton-live --marketplace CanArslanDev/ableton-live-mcp
(Before Claude Code 2.1.275: /plugin marketplace add CanArslanDev/ableton-live-mcp,
then /plugin install ableton-live@ableton-live-mcp.) Besides the MCP server,
the plugin brings five skills Claude uses when they fit: working in Live, mixing
and mastering, composing and arranging, Turkish makams and usuls, and setup and
troubleshooting. Only the MCP server, without the skills:
claude mcp add ableton -- uvx --from git+https://github.com/CanArslanDev/ableton-live-mcp ableton-mcp
{
"mcpServers": {
"AbletonMCP": {
"command": "uvx",
"args": ["--from", "git+https://github.com/CanArslanDev/ableton-live-mcp", "ableton-mcp"]
}
}
}
git clone https://github.com/CanArslanDev/ableton-live-mcp
cd ableton-live-mcp && uv sync
Point the client at uv --directory /path/to/ableton-live-mcp run ableton-mcp,
or load the plugin from disk with claude --plugin-dir /path/to/ableton-live-mcp.
2. Install the Remote Script into Live
Ask the assistant to install the Ableton remote script (the
install_remote_script tool), or run it from a terminal:
uvx --from git+https://github.com/CanArslanDev/ableton-live-mcp ableton-mcp-install-script
This copies the script to <User Library>/Remote Scripts/AbletonMCP
(~/Music/Ableton/User Library on macOS, Documents\Ableton\User Library on
Windows). Restart Live, then choose Settings → Link, Tempo & MIDI → Control
Surface → AbletonMCP, Input and Output None. Restart Live after every
script update.
Run only one MCP client against Live at a time. Ask the assistant to call
doctor to confirm the server and the script
versions match.
Environment variables: ABLETON_MCP_PROFILE=full (every original tool),
ABLETON_HOST / ABLETON_PORT (where the server looks for Live),
ABLETON_MCP_TOKEN (the session token, when the server cannot read
~/.ableton-mcp/live-session.json). In Live's environment:
ABLETON_MCP_VERBOSE=1 (log every command to Live's Log.txt) and
ABLETON_MCP_AUTH=0 (accept clients that do not send the token).
How it works
AI client ──MCP (stdio)──▶ MCP_Server (Python 3.10+)
│ JSON lines over TCP 127.0.0.1:9877,
│ session token from ~/.ableton-mcp
▼
AbletonMCP Remote Script (inside Live)
│ polled on Live's main thread every 10 ms
▼
Live Object Model
The Remote Script runs inside Live's own Python. It polls its socket from
Live's main thread with Live's own timer, so a command is read, run and
answered in about 10 ms (the earlier socket-thread design needed ~400 ms,
because Live only lets background threads run on its 100 ms tick). Besides the original fixed
commands it serves a small generic protocol — lom_get, lom_set, lom_call,
batch, observe/poll_events, extended note and envelope editing, and
setting a parameter by its display value. Most tools are written on the server
side in terms of those, so new Live features rarely need script changes.
Addressing objects
Paths follow the Max for Live convention:
live_set tracks 2 devices 0 parameters 3
live_set tracks "Lead Synth" mixer_device volume # select by name
live_set tracks -1 # last track
live_set return_tracks 0 devices "Reverb"
id 140234 # id from an earlier result
Values that must be Live objects are passed as {"$ref": "<path>"} (for example
a routing type: {"$ref": "live_set tracks 2 available_input_routing_types \"Resampling\""});
enum arguments as {"$enum": "Live.Song.RecordingQuantization", "value": 5}.
Tools take tracks as an index, an exact name, "master" or "return A".
Undo
Live keeps appending changes made by a script to one open undo step until the user touches Live — one Cmd+Z would revert everything the assistant did. The script closes the undo step after every modifying command, so each tool call is its own entry in the undo history. Live still files parameter changes (volume, tempo, mute, device values) separately from structural edits (names, colors, new tracks, clips, notes), so a call that does both can take two Cmd+Z.
Limitations
- Live's API cannot draw track automation, and it refuses envelopes on
Arrangement clips.
write_arrangement_automationbuilds the envelope on a Session clip and lays it over the stretch (MIDI tracks; the copy keeps it);record_automationrecords real track automation, but in real time, out loud. - Live has no export/render API.
render_and_analyze,capture_stemsandmix_doctorrecord in real time on temporary tracks (deleted afterwards; the audio files stay in the project), so they take as long as the section plays. - Arrangement clip creation and
move_deviceneed Live 12;add_deviceandapply_device_chainneed Live 12.3; the extended note API needs Live 11. - Live's API cannot map parameters to rack macros, group tracks, bounce, freeze or export.
- The socket only listens on
127.0.0.1and wants the token Live writes to~/.ableton-mcp/live-session.jsonat every launch (readable only by your user). That keeps other users and web pages out, not other programs running as you.
Troubleshooting
Start with uv run ableton-mcp-doctor in a terminal, or ask the assistant to
run doctor: it checks versions, installed copies of the script, the
connection, open dialogs and Live's log, and says how to fix what it finds.
Development
uv sync
uv run --with pytest pytest -q # 360 tests, no Live or models required
(cd extension && npm ci && npm test) # the Extensions SDK add-on
uvx ruff check .
uv run python packaging/build_mcpb.py # dist/ableton-live-mcp-<version>.mcpb
claude plugin validate . # the Claude Code plugin and its skills
CI (.github/workflows/ci.yml) runs the tests on macOS and Linux with Python
3.10 and 3.13 (Windows reports without blocking), lint, and builds the wheel,
the sdist and the .mcpb. Pushing a v<version> tag publishes them as a GitHub
release.
MCP_Server/fake_live.py is an in-memory imitation of the Live Object Model;
the end-to-end tests run the real Remote Script TCP server and the real MCP
tools against it. See CLAUDE.md for the conventions (the bundled
script copy, version bumps, main-thread rules, Live quirks).
Without Live: the simulator and the eval suite
uv run ableton-mcp-simulator serves that imitation through the real Remote
Script, the way Live does, so the server and its tools run without Ableton
(no sound; the browser holds a few samples). Point a server at it with
ABLETON_MCP_SESSION_FILE=~/.ableton-mcp/simulator-session.json.
evals/ holds 32 claude plugin eval cases (mixing,
composing, makams and usuls, arrangement and transitions, memory, safety) that run the plugin's real
server against a fresh simulated set each time and check the resulting set,
not the reply:
claude plugin eval . --scaffold --allow-real-servers \
--allow-tools "mcp__plugin_ableton-live_ableton__*" --ablation none -j 4
Privacy
Nothing leaves your machine. The server talks only to Live on localhost and reads no credentials. The upstream project's telemetry and dataset recorder (Supabase) were removed in 2.2.0; the Remote Script no longer installs the listeners that fed it.
License
MIT — see LICENSE. Original work © Siddharth Ahuja; extensions © Can Arslan. Not affiliated with or endorsed by Ableton.