Hey everyone,
I built an independent community MCP server for Ultralytics Platform.
It lets MCP clients like Claude Code, Codex, Cursor, VS Code/Copilot, and other agent tools to work with Ultralytics Platform workflows directly from chat/agent sessions.
You can also check the GitHub repo for setup instructions, supported tools, and more detailed examples.
I built it because I wanted AI agents to work with my Ultralytics projects, datasets, training runs, predictions, and exports from the same place I already use them.
Current support includes :
- Browse projects, datasets, models, exports, and GPU availability
- Search public projects and datasets on Ultralytics Explore
- Upload datasets from archives, folders, or videos
- Start dataset ingest jobs
- Monitor training progress and recent metrics
- Run model predictions from image URLs or base64 inputs
- Download model weights
- Create training and export jobs (with explicit cost confirmation)
It’s open source and published:
Would love your feedback 
Nice work @amanharshx, thanks for sharing this with the community. I appreciate that the repo clearly marks this as an independent community project and that billable actions like training/export require explicit confirmation.
One suggestion: keep the safety model very conservative around ULTRALYTICS_API_KEY. Platform API keys are bearer tokens, so it’s worth documenting that users should pass them through MCP client environment config only, avoid committing .mcp.json files with real keys, and review any local file paths before approving upload/download tools. Aligning the README closely with the current Ultralytics Platform REST API docs will also help users handle auth, rate limits, and endpoint changes correctly.
This is a useful community experiment for agent-driven Platform workflows. If users run into Platform API issues while testing it, feel free to open a specific thread here with the request, response status, and minimal repro so we can help debug the Ultralytics side.
1 Like
Thanks, @pderrenger, I really appreciate your feedback.
I’ve tightened the README around API key handling, `.mcp.json` safety, local path review, and redacted Platform API troubleshooting details.
Will keep the safety model conservative as this evolves.
Thanks for the quick update, @amanharshx. That sounds like the right direction.
For future additions, I’d keep the same pattern: fail closed on ambiguous actions, require explicit confirmation for anything billable or destructive, and keep logs