MCP is all you need — Samuel Colvin, Pydantic

Introduction to Pydantic and MCP 00:00

• Samuel Colvin introduces himself as the creator of Pydantic, a popular Python data validation library with 360 million downloads per month.
• Pydantic is widely used in general Python development and GenAI, including SDKs and agent frameworks.
• Colvin's company, Pydantic, has developed Pydantic AI (an agent framework) and Pydantic Logfire (an observability platform).
• He is an inactive co-maintainer of the MCP Python SDK.
• The talk's title is a play on previous talks ("Pydantic is all you need"), emphasizing that MCP can simplify many use cases in agent communication.

MCP Primitives and Tool Calling 02:03

• The focus is on autonomous agents or code, not UI-based coding agents like Cursor or Zed.
• Of MCP's three primitives (prompts, resources, tool calling), prompts and resources are less relevant for this use case.
• Tool calling is highlighted as especially valuable and more complex than just using OpenAPI.
• MCP supports dynamic tools, logging during execution, sampling (a powerful but confusingly named capability), tracing, and running as a subprocess over standard I/O.
• MCP enables agents and tools to be composed agnostically, similar to how browsers and websites interact using open protocols.

Sampling in MCP 05:01

• Sampling allows a remote MCP server to use the client's LLM, reducing cost and complexity.
• The client proxies LLM requests on behalf of the server; this chain can continue through multiple tool calls.
• Sampling enables features like dynamic LLM access but is not widely supported yet.
• Pydantic AI is adding support for MCP sampling as both client and server.

Example: Research Agent with BigQuery Tool 07:10

• A demo agent is configured to research open-source packages using a single tool that queries the BigQuery public dataset for PyPI downloads.
• Code is shown configuring dependencies, retries, and a system prompt that describes table/schema details.
• Output validation includes removing markdown quotes, table name checks, executing queries, and retrying with the LLM if the query fails.
• Logging via MCP context provides progress updates to users before completion.
• This technique allows for interactive updates (logs or progress) while an agent or tool is still working.

Running an MCP Server and Integrating Tools 10:50

• An MCP server is set up using Fast MCP, registering the BigQuery tool for PyPI downloads.
• Tool descriptions (from docstrings) are presented to the LLM for selection.
• Offloading SQL generation and related instructions to the tool (rather than the central agent) keeps the main agent’s context window smaller and more manageable.
• The server runs over standard IO, easily integrating with agent frameworks.

Demo Results and Observability with Logfire 13:01

• The agent can answer questions like “How many downloads did Pydantic have this year?” with real data (e.g., 1.6 billion downloads).
• Observability through Logfire allows inspection of the full workflow, from outer agent’s prompt to tool decision, SQL query generation, and final results.
• The system shows chaining of client-server-client calls, model usage, and returns structured XML-ish query results to the LLM for natural language output.
• Actual SQL generated by the agent can be inspected for correctness and transparency.

Conclusion and Q&A Invitation 15:07

• Samuel invites questions and discussion at the Pydantic booth, welcoming further exploration or demonstrations of potential failure cases.