Pi is an open-source terminal-based AI coding agent with a small extensible core. The project’s October 21 version 1.4.0 makes it a timely candidate for an experiment, especially if you want a more transparent way to connect a coding model to shell commands, project files and custom tools. It is not a replacement for Cursor’s editor; it is a different way to orchestrate coding work.
The design is relevant to a two-person consultancy and to the MiniMac’s collection of small services. TypeScript extensions, skills and custom prompts can turn a simple agent into a repeatable workflow without building an entire new dashboard. The downside is that a terminal agent can also make mistakes quickly if given write or shell access without boundaries.
Compare Pi with Cursor on a disposable copy of a small repository: ask each to add one test and fix one known issue. Measure time, diff quality, approval friction and token costs. Keep production services out of the first trial.
What to do: Run Pi on a throwaway repo and compare one bounded task against Cursor; do not give it production credentials.