Update: the Ultralytics MCP now covers the full model lifecycle (40+ tools)

Discussion
1

Hey again everyone, a big update since I first shared this.

Back then the MCP could browse, train, and predict. Since then it’s grown into the full lifecycle — you can now take a model from an empty project all the way to a live, monitored deployment, entirely from your AI agent.

Here’s a quick end-to-end demo:
I asked Claude to build a tennis-ball detector on the Ultralytics Platform via the MCP — it created the project, uploaded my dataset, checked the class stats, trained a YOLO26 model, showed me the evaluation plots, downloaded the weights, and set up monitoring on the deployed endpoint. All approved step-by-step from chat.

(Couldn’t post a video here, so had to post on youtube)

What’s new since the last post:

  • Deployments & monitoring — list deployments, check health, read logs, pull metrics (requests/latency/errors),
    run live inference, and stop an endpoint.
  • Model evaluation — best-vs-final epoch metrics, plus evaluation plots (PR/F1/precision/recall curves + confusion matrix).
  • Dataset class stats — per-class annotation counts to catch imbalance before training.
  • Multi-dataset fine-tuning — train sequentially across several datasets in one call.
  • More safety rails — cancel running training/export jobs, and an explicit guard before a retrain overwrites a model’s history (keeping the fail-closed pattern from earlier feedback).
  • Reliability — every tool now runs against the live owner-scoped Platform API, with clearer error messages and live smoke tests.

That’s 40+ tools total now, all callable from MCP clients like Claude, Codex, Cursor, VS Code/Copilot, and other agents.

Still open source, same links:

Feedback always welcome :slightly_smiling_face: