AI music agents: run the pipeline
What an agent can actually do with a music pipeline — CLI, MCP and API — and where a human still has to sit in the loop.
An agent that makes music is not a novelty any more. The interesting question is which parts of the job it can hold on its own.
What an agent handles well
Batch production. “Make twenty lo-fi tracks, 90 to 100 bpm, cover art for each, 9:16 visualizer, drop them in a folder.” This is a loop with parameters. Agents are good at loops with parameters.
Format fan-out. One master, then 16:9, 9:16, 1:1 and a Canvas loop from it. Purely mechanical.
Retry and recovery. A render fails, it re-queues with a different model. No human needs to see that happen.
Scheduling. Render overnight, publish on a cadence, respect the rate limits.
What it handles badly
Taste. An agent cannot tell you which of twenty takes is the one. It can tell you which is loudest.
Brief writing. The prompt is the creative decision. Delegating it means delegating the song.
Knowing when to stop. Left alone, an agent will happily generate two hundred tracks nobody asked for and bill you for all of them. Set a spend cap.
How it connects
Three surfaces, same pipeline:
- CLI. For scripts and cron.
vocaler render album.json - MCP. For agents in Claude, or any MCP client. The tools map to pipeline stages.
- REST API — for your own backend.
MCP is the one worth understanding. It exposes each stage — lyrics, song, cover, video, publish — as a callable tool, so an agent can plan across them rather than shelling out blindly.
A workable division
Let the agent own production. Keep the brief and the final pick.
That is not a limitation of the tooling. It is where the value actually sits — the agent removes the tedium, and the tedium was never the interesting part.