Patronus MCP Server

What is Patronus MCP Server?

An MCP server implementation for the Patronus SDK, providing a standardized interface for running LLM system optimizations, evaluations, and experiments.

How to use Patronus MCP Server?

Clone the repository, create a Python virtual environment, install dependencies with uv pip install -e ., then run the server with an API key via --api-key argument or the PATRONUS_API_KEY environment variable. Use the included test script for interactive evaluation or call the server’s tools through any MCP client.

Key features of Patronus MCP Server

• Initialize Patronus with API key and project settings
• Run single evaluations with configurable evaluators
• Run batch evaluations with multiple evaluators
• Run experiments with datasets
• Create custom evaluator criteria in the Patronus API
• List all available evaluators and their criteria
• Custom evaluate using user-defined evaluator functions

Use cases of Patronus MCP Server

• Evaluate LLM outputs for hallucinations or conciseness using remote evaluators
• Run batch evaluations with multiple remote evaluators on the same input
• Conduct experiments with a dataset and a mix of remote and custom evaluators
• Create new evaluator criteria to match specific pass/fail conditions
• Integrate custom evaluation logic via @evaluator decorated functions

FAQ from Patronus MCP Server

How do I supply the Patronus API key?

You can provide it either as a command‑line argument (--api-key your_key) or by setting the PATRONUS_API_KEY environment variable.

What runtime dependencies does the server require?

Python 3, a virtual environment, and the packages installed via uv pip install -e . and uv pip install -e ".[dev]" for development.

Does the server support custom evaluators?

Yes. You can define evaluator functions using the @evaluator decorator and pass them to the custom_evaluate tool, or configure a custom evaluator adapter for experiments.

How can I test the server interactively?

Run python -m tests.test_live src/patronus_mcp/server.py with an API key to get a menu of test options (single evaluation, batch evaluation). Results are shown as formatted JSON.

What transport or authentication does the server use?

The server uses the Model Context Protocol (MCP) and authenticates with the Patronus API via the provided API key. No other transport or auth method is documented.