Pi

// Overview

Pi is an open-source agent harness whose headline piece is a terminal coding agent CLI. The project's stated aim is that you adapt Pi to your workflow rather than the other way around: instead of forking the tool and editing its internals, you extend it with TypeScript extensions, skills, prompt templates and themes, and you can bundle those into "Pi packages" shared over npm or git.

The agent ships deliberately small. Out of the box the model gets four tools — read, write, edit and bash — and features such as sub-agents and a plan mode are left out on purpose; the project's answer is to ask Pi to build the behaviour you want or install a third-party package that already has it. It runs in four shapes: interactive TUI, print/JSON output for scripting, RPC for process integration, and an SDK for embedding in your own app.

The repository is a monorepo, so the parts are reusable on their own: @earendil-works/pi-coding-agent is the CLI, pi-agent-core is the agent runtime with tool calling and state management, pi-ai is a unified multi-provider LLM API, pi-tui is a terminal UI library with differential rendering, and pi-telemetry holds vendor-neutral telemetry contracts. Everything is MIT-licensed TypeScript.

// What it does

• Interactive terminal coding agent with four core tools — read, write, edit and bash
• Extension points in TypeScript: extensions, skills, prompt templates and themes, packaged and shared as Pi packages
• Four run modes — interactive, print/JSON, RPC for process integration, and an embeddable SDK
• Reusable packages: pi-ai (multi-provider LLM API), pi-agent-core (agent runtime), pi-tui (terminal UI), pi-telemetry
• Authenticate with a provider API key or an existing subscription (Anthropic Claude Pro/Max, OpenAI ChatGPT Plus/Pro via Codex, GitHub Copilot) plus many API-key providers
• Supply-chain hardening: exact-pinned direct dependencies, a published shrinkwrap, and installs that use --ignore-scripts
• MCP servers over stdio or streamable HTTP with OAuth, plus a codemode tool that runs model-written JavaScript in a QuickJS sandbox to call tools in parallel (since v0.99.0)
• Full-screen TUI by default and image generation from codemode via models.generateImages() (since v1.0.0)

// Getting started

Install the CLI from npm (or the install script), point it at a provider, and run pi in the project you want to work on.

npm install -g --ignore-scripts @earendil-works/pi-coding-agent

curl -fsSL https://pi.dev/install.sh | sh

export ANTHROPIC_API_KEY=sk-ant-...
pi

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• You want a terminal coding agent whose behaviour you can change in TypeScript without maintaining a fork
• You prefer a small default tool set (read/write/edit/bash) and would rather add capabilities as skills or packages
• You need to drive a coding agent from another process — print/JSON output, RPC, or the embeddable SDK
• You are building your own agent and want the pieces separately: a multi-provider LLM API, an agent runtime, or a terminal UI library

// Version history

Every verified update to Pi that AI/TLDR tracked, newest first — each links to our coverage and the official changeset.

1. 2026-10-01v1.0.0

   First stable release: the full-screen TUI becomes the default, codemode uses about 40% fewer prompt tokens and can generate images, and MCP OAuth adds RFC 9207 iss validation with per-server credentials. Earendil also released Pi Durable, an experimental package for crash-safe, long-running agents.

   our coverage →changeset ↗
2. 2026-09-29v0.99.0

   Adds MCP server support over stdio and streamable HTTP with OAuth, and a codemode tool that runs model-written JavaScript in a QuickJS sandbox to call tools in parallel. Extensions can also register virtual models that route requests, for example with the Jev classifier.

   our coverage →changeset ↗

// How Pi compares

Pi alongside other open-source autonomous coding agents tools AI/TLDR tracks, ranked by GitHub stars.