ableton-connect is an MCP server I am building — it gives conversational AI deep, listen-first access to an Ableton Live session, so an agent can compose, design sound, mix and master alongside me, in context.

End to end, today: the agent analyzes every audible track — key, tempo, density, chord progression — picks an instrument from my own Ableton browser to match the role and the detected genre, loads it, and writes MIDI that follows the actual harmony of the session. 347 tools across session control, audio analysis, perception and workflow planners (EQ, compression, limiting, master readiness, mix audit, reference matching).

It also checks its own work: every plan returns concrete device moves with expected results, measured before and after — and if a change makes things worse, it gets reverted. All audio analysis runs locally; the audio never leaves the machine.

Dialogue over automation: the agent proposes, I decide — accept, reject, or push further. The taste, and the final call, stay mine.

v1 is live-verified against a real Ableton session, with 1080+ tests passing. v2 — real-time listening — is next.