PydanticAI MCP Experiment

Minimalist examples to provide your own MCP servers to your local llm models.

Under examples/ are different configurations providing an MCP client and server using pydantic_ai and local models / ollama to get your hands dirty and tinker.

Setup

This project is uv managed. So to set up run

uv sync

Examples

There are 4 different examples for running a MCP server (locally) via:

• Stdio -> examples/0_subprocess
• HTTP
  • mcp.run -> examples/1_http_mcp_run
  • FastAPI -> examples/2_http_fastapi
  • Starlette -> examples/3_http_starlette

Each folder contains a client.py and server.py file.

The servers should be started before running the client.py file.

The scripts are set up such that when you change into the respective example folder you can run each simply using

python server.py

and

python client.py

Logging

Since pydantic_ai is used here the natural choice for logging is logfire. Hence upon running the client.py files you will be asked to authenticate with logfire. You can change the backend to something like otel-tui, but I tried and failed to find the prompt and response messages I wanted to see.

References

• pydantic + ollama: https://ai.pydantic.dev/models/openai/#example-local-usage
• pydantic ai + mcp + "stdio" server: https://ai.pydantic.dev/mcp/client/#mcp-stdio-server
• logfire setup: https://logfire.pydantic.dev/docs/#logfire - requires account creation
• logfire with otel-tui backend: https://ai.pydantic.dev/logfire/#logfire-with-an-alternative-otel-backend -> does not show the prompts and responses (despite them being present in the screenshots in the docs)