GitHub - Metacreation-Lab/midigpt-LIVE: A multitrack MIDI generation machine for Ableton Live: a Max for Live device powered by MIDI-GPT

MIDI-GPT for Live is a Max for Live device that writes MIDI into your Ableton Live clips with MIDI-GPT, a transformer for computer-assisted multitrack composition. It reads every MIDI clip in the selected scene as context, generates the bars you choose, and writes the result back into the clips as takes you can audition, revert or keep.

Fill in bars: paint the bars to rewrite on any track; the rest of the scene stays as it is and is used as context.
Continue or create tracks: turn on AR to generate a whole track, including one that has no clip yet.
Steer the output: polyphony, note length, density, scale, register, rhythm and remix per track; temperature, top-p and top-k for the whole request.
Takes: generate up to 16 versions at once, step through them in place, go back to the original, or keep one.
Runs on your computer: the model runs in a local server; nothing leaves your machine unless you point the device at a server elsewhere.

Warning

MIDI-GPT for Live is in beta. It works, but it hasn't been used on many machines or sets yet, and some combinations of settings still go wrong. Please report what you find: small things too.

Trailer: watch on YouTube (1:39).

Documentation: metacreation-lab.github.io/midigpt-LIVE: install, a tour of every control, workflows and troubleshooting.

Companion project: MIDI-GPT for REAPER.

Installation

Requirements


Ableton Live	11 or 12: Suite, or Standard with the Max for Live add-on
Max	the Max bundled with Live: 8.6 (Live 11 to 12.1) or 9 (Live 12.2 and later)
OS	macOS, Windows or Linux (tested on macOS; the Windows and Linux installers are untested)
Server	Python 3.10–3.14 (3.10–3.12 on Intel Macs) and `midigpt` 0.4.2; the installer sets these up
Hugging Face	a free account that has accepted the model license, and an access token: see Hugging Face Access
Disk	about 1 GB for the server environment, plus the model: 83 MB (`yellow_small`) or 350 MB (`yellow_medium`)
Hardware	any recent computer; no GPU needed (an NVIDIA GPU is used when there is one)

One-Line Install

macOS and Linux, in Terminal:

curl -fsSL https://metacreation.net/midigpt-LIVE/bootstrap.sh | bash

Windows, in PowerShell:

irm https://metacreation.net/midigpt-LIVE/bootstrap.ps1 | iex

This downloads the latest release and copies the MIDI-GPT folder into your Live User Library (the location is read from Live's preferences). It then asks whether to set up a server on this computer: a Python environment with midigpt (in ~/Library/Application Support/MIDI-GPT/server, ~/.config/MIDI-GPT/server on Linux, %APPDATA%\MIDI-GPT\server on Windows), a Start MIDI-GPT Server launcher (Desktop and Start menu shortcuts on Windows), and the model. On Windows it also offers Microsoft's Visual C++ runtime, which PyTorch needs, and the CUDA build of PyTorch when there is an NVIDIA GPU. If no Hugging Face token is saved yet, it asks for one (see Hugging Face Access). Skip the server if yours runs on another machine.

Options go after bash -s --, for example curl -fsSL …/bootstrap.sh | bash -s -- --no-server:

Flag
`--no-server` / `--server`	install only the devices, or set up the server without asking
`--model yellow_small`	the smaller, faster model (default `yellow_medium`)
`--user-library DIR`	install the devices into another User Library
`--zip FILE`	install from a downloaded release zip
`--yes`	accept every default (a Hugging Face token is then read from `HF_TOKEN`)
`--uninstall`	remove everything (see below)

On Windows the options are PowerShell flags, passed through a script block, for example & ([scriptblock]::Create((irm https://metacreation.net/midigpt-LIVE/bootstrap.ps1))) -NoServer: -NoServer / -Server, -Model, -UserLibrary, -Zip, -Gpu / -Cpu (the PyTorch build), -Yes, -Uninstall.

Manual Install

The device. Download MIDI-GPT-for-Live-<version>.zip from the latest release, unzip it, and move the MIDI-GPT folder into your User Library (by default ~/Music/Ableton/User Library). Keep its contents together:

file
`MIDI-GPT.amxd`	the device you use
`MIDI-GPT Instrument.amxd`	a small tag MIDI-GPT adds to each track to remember its instrument; you never add it yourself
`Server/`	the server launchers: Start MIDI-GPT Server.command (macOS), .bat (Windows), `start-server.sh` (Linux)

A Python environment with Python 3.10–3.14 (3.10–3.12 on an Intel Mac), in the server folder the launchers look in:

OS	server folder
macOS	`~/Library/Application Support/MIDI-GPT/server`
Linux	`~/.config/MIDI-GPT/server`
Windows	`%APPDATA%\MIDI-GPT\server`

python3.12 -m venv ~/Library/Application\ Support/MIDI-GPT/server/.venv     # Windows: py -3.12 -m venv "%APPDATA%\MIDI-GPT\server\.venv"
source ~/Library/Application\ Support/MIDI-GPT/server/.venv/bin/activate    # Windows: "%APPDATA%\MIDI-GPT\server\.venv\Scripts\activate"

MIDI-GPT:
```
pip install "midigpt[http,inference]==0.4.2"
```
On an Intel Mac, add "torch<2.3" "numpy<2" (the last PyTorch built for Intel Macs).
Sign in to Hugging Face with your token (see Hugging Face Access):
```
hf auth login
```
Start the server from the Server folder inside MIDI-GPT: double-click Start MIDI-GPT Server.command (macOS) or Start MIDI-GPT Server.bat (Windows), or run start-server.sh (Linux). The first start downloads the model. Leave the window open while you work; Ctrl+C stops the server. To use another model or port, put MODEL=yellow_small or PORT=3457 lines in a server.conf file in the server folder.

The server uses the GPU when there is one (CUDA, or MPS on a Mac). On a Mac, MPS can be slower than the CPU for a model this size (twice as slow on an Intel Mac with a Radeon GPU): if generations are slow, start it with start-server.sh --device cpu.

Hugging Face Access

The models are downloaded from Metacreation/MIDI-GPT on Hugging Face, which needs a free account:

Sign in at huggingface.co and accept the license on the model page.
Create an access token in Settings › Access Tokens (huggingface.co/settings/tokens); the Read type is enough.
Give it to the installer when it asks, or run hf auth login in the server's environment. The token is saved in ~/.cache/huggingface/token, where the server finds it on every start. Setting HF_TOKEN works too.

The token is only used by the server, to download the model; the device never sees it. A model that is already downloaded keeps working without one. If the server can't get the model, its window says why and stays open.

macOS Quarantine

macOS quarantines files downloaded with a browser. If Live won't load the device, or the launcher won't open, clear the flag (the one-line installer does this for you):

xattr -dr com.apple.quarantine ~/Music/Ableton/User\ Library/MIDI-GPT

Updating / Uninstalling

Run the one-line install again to update: it replaces the devices and keeps your settings and server. To remove everything:

curl -fsSL https://metacreation.net/midigpt-LIVE/bootstrap.sh | bash -s -- --uninstall

On Windows, in PowerShell:

& ([scriptblock]::Create((irm https://metacreation.net/midigpt-LIVE/bootstrap.ps1))) -Uninstall

This removes the MIDI-GPT folder from the User Library, the server environment, the launchers and the settings. Downloaded models stay in ~/.cache/huggingface (delete models--Metacreation--MIDI-GPT there to free the space), and so does the Hugging Face token (hf auth logout removes it).

How It Works

 Live set ─── clips in the selected scene ───▶ MIDI-GPT.amxd ─── HTTP ───▶ midigpt-http (127.0.0.1:3456)
    ▲                                               │                           │  the MIDI-GPT model
    └─────────── generated bars, as takes ◀─────────┘ ◀──────── score ──────────┘

The device sends the scene to the server as a score, with what to generate and how. The server runs MIDI-GPT and returns the result, which the device writes into only the bars you asked for. The server is the HTTP server from the midigpt Python package (API reference). The device keeps the server address in ~/Library/Application Support/MIDI-GPT/live.json.

Usage

Quick Start

Start the server: Start MIDI-GPT Server on the Desktop or in the MIDI-GPT folder's Server folder (step 5 above). On the device's SETUP page, the light next to the address turns green when the server is reachable.
In Live's browser, open User Library › MIDI-GPT and drag MIDI-GPT onto any track. Click the key on the device to open its window.
Run SETUP › SETUP ALL once per set: MIDI-GPT guesses each track's instrument from its device and name, lets you confirm each guess, and writes a small MIDI-GPT Instrument tag onto each track.
Select a scene with some MIDI clips. In the roll at the top of the window, paint the bars to generate on a track, or turn on AR for the whole track.
Press GENERATE. Step through the takes, then press KEEP.

Choosing What to Generate

The roll shows every MIDI track of the selected scene. A track is generated when you paint some of its bars or turn on AR; every other track is context. Mark a track ignore to leave it out entirely. A track with no clip in the scene gets one when it is generated. All clips in the scene must share one time signature.

Takes

Set TAKES above 1 to get several versions from one request. The OUTPUT section lists them: click a take to write it into the clips, or the ↩ key to hear what was there before. KEEP commits the take you are on. The takes stay while you work elsewhere in Live, and always go into the clips they were generated for.

Track Instruments

MIDI-GPT needs to know what each track plays: a General MIDI instrument, or drums. That is stored on the track itself, in the MIDI-GPT Instrument tag, so it survives renaming, moving and duplicating tracks, and is saved with the set. Change it from the tag's FAMILY and PROGRAM menus or from SETUP. A track without a tag is guessed from its instrument (a Drum Rack means drums), then from its name, else piano.

The device does not play anything: your tracks keep their own instruments and sounds.

Remote Server

The server can run on another computer, such as a studio machine with more power. Start it there with --host 0.0.0.0 (start-server.sh --host 0.0.0.0) so other machines can reach it, then enter its address (for example http://192.168.1.20:3456) on the device's SETUP page with EDIT. The address is remembered for every set.

Controls

TRACK, for the selected track. Each knob has an AUTO setting that leaves it to the model.

Control	What it does
POLY	how many notes sound at once (melodic tracks)
LENGTH	the range of note lengths, from 32nd to whole notes (melodic tracks)
DENSITY	how busy the part is (drum tracks)
KEY · RANGE · SILENCE · PCS	prism and expressive: key, pitch range, silent share of each bar, pitch classes per bar
EXPRESSIVENESS	expressive: how loosely the notes sit on the beat, from on the grid to freely timed
GENRE	prism and expressive: the style to write in, for the whole request; click for the list, or ◀ ▶

SAMPLE, for the whole request.

Control	Range (default)	What it does
TEMP	0.1–3 (1)	randomness: lower is safer and more repetitive, higher is more adventurous
TOP-P	0.05–1 (off)	sample only from the most likely choices that add up to this probability
TOP-K	0–500 (off)	sample only from the K most likely choices
ESC	1–3 (1)	raise the temperature on each retry after a rejected result
TAKES	1–16 (1)	how many versions to generate
PROB	5–100 % (ALL)	regenerate only this share of the painted bars, picked at random
SEED	1–2147483647 (random)	click the screen to type one; DICE picks a random seed, LOCK keeps the last run's (or the selected take's). A fixed seed repeats a result exactly.
NOVELTY · SILENCE · SHUFFLE	on / off	reject copies of the context, reject empty bars, vary the order tracks are generated in

MASKS, for the selected track.

Control	What it does
SCALE · ROOT · MODE · REGISTER	keep notes in a key and between two pitches
RHYTHM · UNIT · GRID / EXACT	guide onsets toward a grid, or place them exactly where you draw them
REMIX · AMOUNT · P+D	vary the existing notes instead of writing new ones: AMOUNT is the share of the notes that change (at least one), pitch only or pitch and duration

The ENGINE page holds the model choice; CONTEXT (how many bars the model sees at once); BARS/STEP and TRK/STEP (how much is generated per step); ATTEMPTS; polyphony and density hard limits; MASK P / K; and VELOCITY and MICROtiming when the model supports them.

When a request fails, the readout under the takes says what went wrong and what to change.

Tips & Common Gotchas

DENSITY only applies to drum tracks; POLY and LENGTH only to melodic tracks.
REGISTER only applies with SCALE on.
REMIX replaces the rhythm settings on the same track. POLY LIM and DENS LIM win over it: a wider chord or a fuller bar comes back with fewer notes.
The scene needs at least CONTEXT bars; lower it for short clips.
The first generation after the server starts is slower while the model loads.
On older computers, yellow_small is noticeably faster.

Troubleshooting

Symptom	Fix
NO LINK in the readout, red light in SETUP	the server isn't running or the address is wrong: start it, or check the address. If it stopped at startup, its window says why
The server window says the model is gated	accept the license at huggingface.co/Metacreation/MIDI-GPT and save a token: see Hugging Face Access
Generation is slow on a Mac	the GPU (MPS) can be slower than the CPU: start the server with `start-server.sh --device cpu`
Windows: the server stops with a DLL error (`WinError 126`, `c10.dll`)	PyTorch needs Microsoft's Visual C++ runtime: install vc_redist.x64.exe
Windows: "Windows protected your PC" when starting the server	click More info › Run anyway, or right-click the `.bat` › Properties › Unblock (the installer unblocks the files itself)
Port 3456 already in use	another MIDI-GPT server (MIDI-GPT for REAPER uses the same port) is already running: use it, or start this one with `--port 3457` and change the address
TAG FAILED	Live didn't load the tag device: drag MIDI-GPT Instrument from the MIDI-GPT folder onto the track
TIME SIG	the scene's clips use different time signatures
TOO FEW BARS	the scene is shorter than CONTEXT: lower it or lengthen the clips
The device won't load after a download	see macOS Quarantine

Settings: ~/Library/Application Support/MIDI-GPT/live.json. Server: ~/Library/Application Support/MIDI-GPT/server (its window shows the log). Models and the Hugging Face token: ~/.cache/huggingface.

Reporting Bugs

MIDI-GPT for Live is in beta, so reports are the most useful thing you can send. Open a bug report: the form asks for what you did, what the OUTPUT screen said, the settings involved, the lines around the error in the server window, and your versions. Reporting bugs says where to find each of them.

When a generation fails, the server saves the request it got to failed_requests.jsonl in its folder (~/Library/Application Support/MIDI-GPT/server on macOS, ~/.config/MIDI-GPT/server on Linux, %APPDATA%\MIDI-GPT\server on Windows). Attaching it lets the failure be replayed exactly, but it contains the MIDI of your clips: only attach it if you're happy to share that.

Contribution

The devices are built from the sources in this repository; CLAUDE.md describes the architecture, the dev bridge and the tools.

patchers/        MIDI-GPT.amxd and MIDI-GPT Instrument.amxd (built by the scripts below)
javascript/      midigpt.js (Live API, state, HTTP client) and the panel, strip, tag and Node helper scripts
media/panel/     the panel's images, rendered in Blender
tools/           build, release, freezing, asset and dev scripts
install.sh       the installer for macOS and Linux (bootstrap.sh downloads and runs it)
install.ps1      the installer for Windows (bootstrap.ps1 downloads and runs it)
server/          the server launchers, shipped as MIDI-GPT/Server: start-server.sh, Start MIDI-GPT Server.command / .bat

python3 tools/build_device.py           # patchers/MIDI-GPT.amxd: panel window, strip key, dependencies (in place)
python3 tools/build_tag_device.py       # patchers/MIDI-GPT Instrument.amxd
python3 tools/export_panel_assets.py    # javascript/midigpt_panel_spec.js (and lo-fi panel images: see CLAUDE.md, Panel assets)
python3 tools/build_release.py          # dist/MIDI-GPT-for-Live-<VERSION>.zip, both devices frozen

build_release.py freezes the devices itself (tools/amxd_freeze.py, the same format as Max's Freeze Device), so a release needs no Max. Attach the zip to a GitHub release tagged v<VERSION>; the installer downloads the newest one. Device conventions follow Ableton's Max for Live production guidelines.

Issues and pull requests are welcome (reporting bugs). Known gaps: a MIDI-mappable Generate control, and testing on Windows and Linux.

Citing

If you use MIDI-GPT in your work, please cite the paper:

@inproceedings{pasquier2025midigpt,
  title     = {MIDI-GPT: A Controllable Generative Model for Computer-Assisted Multitrack Music Composition},
  author    = {Pasquier, Philippe and Ens, Jeff and Fradet, Nathan and Triana, Paul and Rizzotti, Davide and Rolland, Jean-Baptiste and Safi, Maryam},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  volume    = {39},
  year      = {2025},
  url       = {https://arxiv.org/abs/2501.17011}
}

To cite the device itself, use this repository's CITATION.cff (GitHub's Cite this repository button).

Acknowledgments

MIDI-GPT for Live grew out of MMM4LIVE, started by Jeff Ens and Pia Baltazar.

Navigation Menu

Repository navigation

About

Repository files navigation

Table of Contents

Installation

Requirements

One-Line Install

Manual Install

Hugging Face Access

macOS Quarantine

Updating / Uninstalling

How It Works

Usage

Quick Start

Choosing What to Generate

Takes

Track Instruments

Remote Server

Controls

Tips & Common Gotchas

Troubleshooting

Reporting Bugs

Contribution

Citing

Acknowledgments

License

About

Releases

Contributors

Languages

Name		Name	Last commit message	Last commit date
.github		.github
docs		docs
javascript		javascript
media/panel		media/panel
patchers		patchers
server		server
tools		tools
.gitattributes		.gitattributes
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
CITATION.cff		CITATION.cff
CLAUDE.md		CLAUDE.md
LICENSE		LICENSE
README.md		README.md
VERSION		VERSION
bootstrap.ps1		bootstrap.ps1
bootstrap.sh		bootstrap.sh
install.ps1		install.ps1
install.sh		install.sh
mkdocs.yml		mkdocs.yml

Latest commit

History

Folders and files

Repository files navigation

Table of Contents

Installation

Requirements

One-Line Install

Manual Install

Hugging Face Access

macOS Quarantine

Updating / Uninstalling

How It Works

Usage

Quick Start

Choosing What to Generate

Takes

Track Instruments

Remote Server

Controls

Tips & Common Gotchas

Troubleshooting

Reporting Bugs

Contribution

Citing

Acknowledgments

License

About

Resources

Stars

Watchers

Forks

Releases

Contributors

Languages