Pi.dev

Pi is an extensible AI agent that works in a terminal. Give it a goal and a working folder, and it can inspect files, run commands, edit content, and carry out several steps. The Pi documentation presents it for software development, research notes, writing projects, and data files.

Work in a folder and continue sessions

Pi supports an interactive workflow in which you add files, direct ongoing work, and export results. You can continue or branch a session while retaining its history. Branching is useful when you want to investigate another approach without discarding the work already recorded in the conversation.

The documentation describes several ways to connect a model: a subscription, an API key, a local model, or a compatible endpoint. Its overview does not give a price or establish that all providers have identical access terms. Choose a connection that matches the work and confirm its costs separately.

Extend or automate the agent

Pi can reuse prompts and load specialized instructions. It also supports executable integrations, terminal interface changes, connections to model services, and packages for distributing those resources. These mechanisms let you adapt an existing agent workflow or build an application around it.

For automation, the documentation distinguishes print mode for one-off scripted tasks, a JSON event stream for structured events, and RPC mode for controlling a separate Pi process. The TypeScript SDK lets a developer run Pi inside an application. The overview identifies these options but does not supply the detailed setup for each one; follow the relevant reference before depending on a particular mode.

Understand the permission boundary

Pi's tools and extensions run with the permissions of the Pi process. Project trust determines which project resources load, but it does not sandbox tool calls. That distinction matters when selecting a working directory or adding an extension: approving project resources does not create a separate execution boundary.

The publisher directs readers to its security guidance before using untrusted files, repositories, extensions, or unattended automation. Start with a folder and permissions appropriate to the task, and review any executable integration before giving it access to the same environment.