Pydantic AI

Pydantic AI is an open-source Python agent framework from the team behind Pydantic, the validation library that already underpins much of the Python AI ecosystem (including the OpenAI and Anthropic SDKs). Its central idea is to bring that same type-safety and structured validation to building agents: agent outputs, tool call arguments, and injected dependencies are defined with Python type hints and Pydantic BaseModel schemas, so errors surface at write time in the IDE rather than at runtime, and the model's responses are validated and, where needed, automatically re-prompted for correction.

The framework is model-agnostic, with a consistent interface across OpenAI, Anthropic, Gemini, DeepSeek, Mistral, Grok, Cohere, and providers reached through Bedrock, Azure, Google Cloud, and Ollama. It includes a typed dependency-injection system for passing data and services into agents, streaming with validated partial outputs, built-in support for Model Context Protocol (MCP) servers, and support for agent-to-UI protocols such as AG-UI and the Vercel AI Data Stream Protocol. A durable-execution layer lets long-running agents checkpoint and resume.

Pydantic AI is free and open source under the MIT license. Its natural companion is Pydantic Logfire, the team's OpenTelemetry-based observability platform, which integrates in a few lines to trace agent runs, tool calls, and token usage — a separately priced product with its own free tier. The framework itself carries no usage fees and is self-hostable anywhere Python runs.