Last updated: 13 August 2026
These terms cover the data AbletonMCP collects. The software itself is licensed under MIT — that license governs the code and grants no rights over your data. This document covers the data.
Running your own fork with your own Supabase credentials? None of this applies. Nothing reaches us, and you become responsible for your own users' data.
AbletonMCP is an independent open-source project maintained by Siddharth Ahuja. There's no company behind it.
Both tiers write to a Supabase project (PostgreSQL) over HTTPS. The anon key shipped in client configs is insert-only and can't read rows back; reads require a service_role key kept in the maintainer's environment.
Contact, including deletion requests: ahujasid@gmail.com
| Anonymous telemetry | Dataset recording | |
|---|---|---|
| Default | On | On |
| Prompts / MIDI | Never | Yes |
| Device & sound design parameters | Never | Yes |
| Toggle | ABLETON_MCP_DISABLE_TELEMETRY=true |
ABLETON_MCP_DISABLE_DATASET=true |
Neither sends anything unless Supabase credentials are configured. No credentials ship with the package.
Collects: a random installation ID, a per-run session ID, package/Python/OS/Ableton versions, which tools ran, success and duration, and error messages with emails and filesystem paths stripped.
Collects no prompts, no musical content, no names, no audio.
The installation ID is a random UUID generated on first run and stored in customer_uuid.txt under your OS app-data directory (~/Library/Application Support on macOS, %APPDATA% on Windows, ~/.local/share on Linux). It isn't derived from your hardware, username, or any account — it exists to count distinct installs rather than raw events. Delete that file and you get a new identity, unlinked from anything recorded before.
Legal basis: legitimate interest in maintaining the software. Opt out any time with the variable above.
On unless you turn it off. Collects your prompts, your MIDI (pitch, timing, duration, velocity), session structure, track and clip names, preference labels, and browser auditions. No audio is ever recorded or uploaded.
It also records sound design state: the full parameter set of every device on every track — including devices nested inside instrument, drum, and audio effect racks, and on the master chain. In practice that means envelope settings (attack, decay, sustain, release), filter and LFO settings, oscillator and synth knob positions, effect parameters, macro values, automation state, and per-clip gain, pitch, and warp settings. If you built a patch or dialled in a mix, the resulting parameter values are recorded alongside the action that produced them.
Generic musical labels (Bass, Drums, Verse) are kept for training signal; other names — including rack chain names — become placeholders like <name:17>. Emails and absolute paths are stripped from prompts and errors. This is best-effort pattern matching, not a guarantee — it can't catch personal information typed into a prompt in an unanticipated form.
Recording is on by default and starts with your first tool call. You are notified once — as a dialog if your client supports it, otherwise as a message in the chat — and can decline there. Recording continues until you do: an unanswered question does not stop it, and on clients that cannot show the prompt you may never be asked at all. Your answer is stored locally in ~/.ableton-mcp/consent.json.
If you are in a jurisdiction where processing this data requires opt-in consent (for example the GDPR, where an opt-out default is generally not a valid legal basis for this kind of personal data), turn recording off with ABLETON_MCP_DISABLE_DATASET=1 before your first tool call.
Withdraw consent any time by saying so in the chat, deleting ~/.ableton-mcp/consent.json, or setting ABLETON_MCP_DISABLE_DATASET=1, which overrides any stored answer.
A non-exclusive, irrevocable, worldwide, royalty-free license to use your recorded trajectories — prompts, MIDI, session structure, device and sound design parameters, preference labels — to train and evaluate models, and to publish or share datasets derived from them.
Derived datasets may be released publicly or shared with research collaborators. That decision hasn't been made yet; the grant is written to allow it.
You keep ownership and copyright in your music. This grants use, not exclusivity. Nothing here limits what you do with your own work.
Only leave this on if you have the right to grant that for everything you record. If you're working on someone else's material, under an NDA, or on a label deal with delivery restrictions, turn it off with ABLETON_MCP_DISABLE_DATASET=1 before your first tool call — it is on until you do.
Rows are kept indefinitely. There's no automatic expiry.
To have yours deleted, email ahujasid@gmail.com with your installation ID (from customer_uuid.txt) or your session IDs. No charge, no need to explain why.
One limit, stated plainly: data already folded into a trained model can't be pulled back out. Training isn't reversible, and retraining from scratch to exclude one contributor isn't something this project can commit to. Deletion covers the stored rows and any future training run — not a model that has already seen them.
If that's not acceptable to you, turn dataset recording off before you start. Deleting afterwards won't fully undo it.
Backups and export snapshots taken before a deletion request may persist until they age out of ordinary rotation.
Depending on where you live (GDPR in the EEA/UK, CCPA/CPRA in California, and comparable laws elsewhere) you may have the right to access, correct, delete, export, or object to processing of your data, and to withdraw consent. Email ahujasid@gmail.com to exercise any of them.
Not directed at children under 13 (under 16 in the EEA). We don't knowingly collect their data — email us if you believe we have, and it'll be deleted.
Material changes get a new date above and a note in the release notes. This is a versioned package, so the terms in effect for you are the ones shipped with the version you're running. If you disagree with a change, disable the relevant tier or stop using the package.
- Anonymous telemetry: on, no creative content, one variable away from off.
- Dataset recording: on, includes your prompts and MIDI, one variable away from off.
- Nothing is sent without configured credentials.
- You can get your rows deleted; you can't un-train a model.