pcode: an open-source terminal coding agent built on Pydantic AI
I built pcode, an MIT-licensed terminal coding agent on Pydantic AI's Harness coder.
The framework-facing part is the extension model: Python extension files add tools, slash commands, guardrails and sub-agents via Pydantic AI capabilities. The repository is a concrete example of using those capabilities in an interactive coding application rather than a standalone agent script.
The terminal layer keeps native scrollback: you can change the visibility of commands, output and diffs after completion without rerunning them, while retaining normal terminal scrolling, search and copy. Long-running commands become background jobs that wake the agent when they finish. Conversations can be rewound and branched, and parallel sessions can use separate git worktrees.
It supports browser automation with your existing logins and models through API keys as well as existing Claude or ChatGPT subscriptions. The agent can edit files and run shell commands with your permissions and no approval prompt, so review the documented tool permissions before using it on a repository.
Source and implementation details: https://github.com/cruxwell/pcode