Ultralytics Platform MCP

What Ultralytics Platform MCP MCP server does

Ultralytics Platform MCP MCP server exposes Ultralytics Platform workflows to an MCP-compatible client. It is intended for operations involving projects, datasets, models, training, prediction, exports, and dataset uploads. An agent can inspect available projects and datasets, create a private project, upload data, start fine-tuning, review training metrics, run a model against a URL or local file, and download model weights.

Projects and datasets are private when created through the server. Project and dataset deletion uses the platform trash flow, with restoration available for 30 days. Dataset deletion moves its images and annotations with it, while models trained on that dataset remain unaffected.

How it works

The server runs as a Node.js stdio process and authenticates requests with an Ultralytics Platform API key. A typical MCP client starts it with npx -y ultralytics-mcp@latest. The API base defaults to https://platform.ultralytics.com/api, with an optional environment variable for advanced endpoint overrides.

Local file operations happen on the MCP client host rather than on a remote server. This applies to local dataset uploads, prediction calls that use file_path, and model downloads. Upload and download URLs are signed and do not forward the Authorization header.

Setup and configuration

Ultralytics Platform MCP MCP server requires Node.js 20 or newer and an API key created in the Ultralytics Platform settings. Put the key in the MCP client's environment configuration, not in prompts, scripts, or committed files. The required variable is ULTRALYTICS_API_KEY; ULTRALYTICS_API_BASE can override the default API endpoint.

For local video dataset uploads, install ffmpeg and ffprobe and make both available on PATH. Configure the server using the client's JSON stdio server settings, then restart the client or session after installation or package upgrades. The repository documents setup examples for Claude Code, Claude Desktop, Codex, Cursor, Gemini CLI, and VS Code/Copilot.

Tools and capabilities

The available workflow areas include:

• Listing and managing Ultralytics projects and datasets.
• Uploading datasets, including local video input when the required media tools are installed.
• Starting training and checking training progress and epoch metrics.
• Running model or deployment predictions against URLs or local files.
• Creating and cancelling exports, then downloading model artifacts.
• Deleting models and moving projects or datasets to trash.

Cost-sensitive actions require confirmation. training_start and export_create require confirm_cost: true. Restarting training on a model with an existing recorded run additionally requires confirm_history_loss: true, because its prior status, epoch count, and training result history are replaced.

Limitations and notes

Ultralytics Platform MCP MCP server reports estimated cost and remaining balance after a training job or export starts; it does not provide those values before the billable operation begins. Cancelling training preserves the latest checkpoint and model. Export cancellation is status-dependent, and a race between status checking and cancellation can remove an artifact if the export finishes first.

Ambiguous project or dataset references fail rather than being guessed, and undeclared tool arguments are rejected. Starting checkpoint-mode training can create a model before the platform validates checkpoint and dataset task compatibility; if validation fails, the created model may need manual deletion. Review local paths before approving uploads, predictions, or downloads, and check download overwrite settings.