fiftyone

The FiftyOne MCP Server exposes FiftyOne's operator framework and plugin ecosystem through the Model Context Protocol (MCP), enabling AI agents to manage datasets, execute operators, and orchestrate computer vision workflows via natural language. It ships with a curated set of built-in tools (dataset management, operator execution, plugin discovery/installation, and session control) and can auto-discover and install FiftyOne plugins as needed by your prompts. In practice, you can instruct an agent to list datasets, load a dataset, run an operator on samples, or install plugins to extend capabilities, all through MCP tool prompts like: “List all my datasets”, “Load quickstart dataset and show summary”, or “Find similar images in my dataset”. The server keeps the operational state and context, enabling fluid multi-step interactions with FiftyOne resources and workflows.

Prerequisites:

• Python 3.8+ (tested with modern Python environments)
• Network access to install Python packages from PyPI

Installation steps:

1. Create and activate a virtual environment (optional but recommended):

python -m venv venv
# macOS/Linux
source venv/bin/activate
# Windows
venv\Scripts\activate

2. Install the MCP server package from PyPI:

pip install fiftyone-mcp-server

3. Run the MCP server locally (example using the module entry point as shown in the README):

# If using a virtual environment, ensure it's activated
python -m fiftyone_mcp_server

4. (Optional) Use Poetry for development workflows as described in the repository if you prefer:

poetry install
poetry run fiftyone-mcp

5. Verify the server is running by issuing a small MCP request or checking logs for a listening port (default behavior may expose HTTP/WebSocket endpoints depending on configuration).

Tips and considerations:

• If you run in a virtual environment or container, keep the environment consistent between the MCP server and your MCP-enabled tools.
• The MCP server integrates with FiftyOne operators (80+ built-ins) and plugins; you can install plugins on-demand when the agent requests capabilities not present initially.
• When configuring clients (Claude, Cursor, VS Code, etc.), you typically point them to the server with a command like fiftyone-mcp or the Python module invocation shown above. The exact configuration file snippets shown in the README (under various tools) demonstrate how to wire up the server name to the command in different environments.
• If you need to upgrade, re-run the pip install fiftyone-mcp-server to fetch the latest release, or use your environment’s package manager as appropriate.
• For debugging, inspect server logs for MCP discovery, operator invocation, and plugin installation events to identify where prompts map to specific FiftyOne actions.