Pi Explained: The Open-Source AI Coding Agent You Can Customize

Coding agents are quietly changing the division of labor in software development.

A growing amount of the mechanical work can now be handed to an agent: inspecting a repository, tracing a bug, editing several files, running tests, or attempting a refactor. The developer increasingly spends more time describing the task, steering the implementation, and checking whether the result is actually correct.

That raises another question. If an agent is going to do more of the work, how much control should you have over the agent itself?

Which model does it use? What tools can it call? What instructions does it load? What happens before a shell command runs? How does it remember a session? And what if you want the same agent loop inside your own IDE or product?

Pi is built around those questions.

So, what exactly is Pi?#

Pi describes itself as a "minimal terminal coding harness." That wording matters more than calling it another AI coding assistant.

Run Pi inside a repository, and you already have a usable coding agent. Its default tools let the model read files, create or overwrite them, patch existing files, and execute shell commands. Additional read-oriented tools such as grep, find, and ls can also be enabled.

The terminal experience includes things you would expect from a modern coding agent: file references with @, model switching, persistent sessions, conversation branching, context compaction, shell commands, and session export.

But Pi's main idea is not that it can edit code. Plenty of tools can do that.