listen_ab comparing before and after is the honest core — the model can tell whether an edit actually did what was asked instead of assuming it did. And the taste tools learn from A/B choices you made, not a universal quality score, which is a more defensible way to remember your preferences. Edits inside the set ride Ableton's undo stack; library and installer changes do not, so those are the ones to watch.
A server that gives an MCP client control over Ableton Live 12: explicit tools for the set, plus library search, listening analysis and taste learning from the versions you keep.
- Direct control over transport, tracks, clips, notes, devices, routing, mixing, automation and Arrangement View
- Library search across Ableton's library, your plugins and racks, local samples and Splice before loading — atlas_search, scan_full_library, augment_with_samples
- Listening before judging — listen_capture reads the result and listen_ab compares it against the state before the edit
- Taste learning from real choices — taste_record_pair, taste_train and taste_rank learn from audio you kept versus discarded
- Composition and device helpers such as compose, insert_device and apply_semantic_move, which runs a named move rather than a tool of its own
- Changes inside the set use Ableton's undo stack; downloads, library scans and installer changes happen outside it and cannot be undone from Ableton
Ableton Live 12 on macOS or Windows. The setup wizard, npx livepilot --setup, installs the Ableton Remote Script, creates the Python environment, copies the Max for Live Analyzer, and checks the connection; Claude Desktop users can double-click a livepilot-<version>.mcpb from Releases. The npm package is livepilot. Licensed BSL 1.1; pin a version for repeatable setups.
One command — npx livepilot