pydantic-deepagents

Why we included this project

Developers who want a Claude Code style workflow without vendor lock get a self-hosted terminal assistant that runs on whatever model they choose, with planning, file edits, shell access, web search, and MCP connections ready to go. The same machinery is exposed as a Python library, so instead of hand-building the tool-calling and sandboxing layers yourself, you can spin up a full agent from one function call and reshape it into your own coding or research tool. A notable feature is live run forking: an in-flight task splits into isolated branches that a judge or a real test command later merges back together, which matters when a refactor could go several directions and you want evidence before picking one. Because execution can be sandboxed in Docker and sessions checkpoint and persist, teams can keep agentic work contained and reproducible rather than merely demo-able.