What is pi-fleet?

pi-fleet is a command line tool for running a small team of restricted, single purpose AI coding agents built on top of the Pi coding agent. Instead of one general purpose assistant, it gives each role, such as implementer, reviewer, researcher, or designer, its own reusable agent seat with a default model, a defined skill, and a list of tools it is allowed to use. A reviewer seat, for example, has no ability to write, edit, or run shell commands at all, while a builder seat does, and that restriction is enforced by the tool rather than just suggested in a prompt. These seats are organized into a small hierarchy: a human still sets goals and priorities, a conductor agent routes work across projects, project leads own a single project and cast worker agents to do the actual implementation, review, or research, and each worker reports back up the chain. This structure lets several agents work on different parts of a project at the same time, each running in its own terminal pane, while keeping real limits on what any one of them can touch. The project does not reimplement any of this from scratch. It combines three existing packages: one that builds the agent's configuration and launches it, one that lets agents spawn other agents, and one that manages fine grained, per command permissions. pi-fleet itself is a thin layer of versioned configuration, wrapper scripts, and skills sitting on top of those three pieces. Setup is done through a single install script that checks for and, on a Mac with Homebrew, offers to install the required dependencies, wires up local configuration without overwriting anything that already exists, and walks the user through authenticating with GitHub, Linear, and their chosen AI provider. It is written in TypeScript and requires Node.js 20 or newer, along with the Pi agent itself already installed and authenticated.