Experiments with Ableton-MCP

Summary

Experiments with ableton-mcp (Dec 2025) I started tinkering with Ableton and MCPs over the holiday break and ended up creating 70+ automation tool calls and a mashup track.soundcloud.com/jhurliman/octo-gloFinding ableton-mcpWhile looking into automation options for Ableton, I stumbled on ahujasid/ableton-mcp. It’s an MCP server that bridges tool-calling LLMs to Ableton Live via a community-documented Python API. I downloaded Claude Code, switched to Opus 4.5, asked it to install AbletonMCP, and then fired up Ableton.Out of the box, the tool set was enough for basic creation and editing in Session view. But it didn’t cover Arrangement view, devices and device chains, mixing, and a bunch of other “real DAW” workflows.Extending ableton-mcpThe fun part: modern LLMs are capable enough to look up docs, add new MCP tools, test them, and iterate in a mostly closed loop. For Ableton features that weren’t exposed by the Python API, Opus 4.5 managed to reverse engineer enough of the .als file format to inject tempo/volume automation and warp markers.Once the low-level plumbing was in place, I tried building higher-level tools. One example was vocal_to_midi(): it analyzes the audio of a vocal track, categorizes onsets into rough phoneme classes, and maps those phoneme categories onto standard Drum Rack MIDI notes. The goal wasn’t to “make drums from vocals” so much as to get a structured representation that helps with tiny phase and timing adjustments when aligning vocals to a drum groove (manually or with automation).Giving the LLM earsEven with better tools, most attempts were still one-shot, or they required a human tightly in the loop. The core limitation was obvious: the model couldn’t hear what it was doing.To address that, I created a Max4Live patch that provided a simple WAV file recorder that could be toggled on and off. This seemed to be the only programmatic way to get audio out of Ableton. Next I deployed two Replicate endpoints:I haven’t gone deep testing closed lo...