Pydantic Deep Agents 🤖 Agent Open source

What it does

Pydantic Deep Agents is two things in one repository: a terminal coding assistant in the style of Claude Code, and the Python framework that powers it. Both are built on Pydantic AI and can use Claude, GPT, Gemini or local models. You can use the CLI directly to plan, edit files, run commands and connect MCP servers, or call a single function in Python to get an agent with the same harness for your own tools and products.

What is inside

The harness provides filesystem and shell tools, planning, persistent memory, skills loaded from SKILL.md files, subagents and teams, context management, MCP client support and optional Docker sandboxing. Its most distinctive feature is live run forking: an in-progress run can split into several isolated branches, each with its own copy-on-write file overlay, steering message and budget cap, and a judge, a vote, a manual choice or the exit code of a test command decides which branch to keep. The CLI adds commands for forking and merging, skills, MCP servers, model switching and compaction, plus a headless runner for scripts. Structured output is type-checked through Pydantic.

Works with

It is a standalone agent and framework rather than an add-on for another coding app. It can import MCP server definitions from a Claude Code setup and reads the same skill format.

How to install

For the framework:

pip install pydantic-deep

For the terminal assistant on Windows or without the install script:

pip install "pydantic-deep[cli]"

Maintenance and safety

MIT licensed, frequently released on PyPI, with CI, coverage reporting and a published security policy. The quick-start also offers a curl-to-bash script from the project's own GitHub repository; the pip route is the more auditable option. The comparison table in the README is written by the maintainers, so treat its claims about other tools as their view. Running forks multiplies model spend, so set budgets.

Who should use it

Python developers who want an open, model-agnostic coding agent they can extend, or who are building their own agents and want planning, sandboxing and memory without writing the plumbing.