Pi Coding Agent

Pi is a minimal, extensible agent harness from Earendil Inc. and contributors, distributed as the @earendil-works/pi-coding-agent CLI under the MIT License. The homepage announces that Pi 1.0 has shipped alongside Pi Durable, and the GitHub repository lists v1.0.0 as the latest release, published 2026-10-01.

The project positions itself as a harness you adapt to your workflow rather than a sealed product. The earendil.com Pi 1.0 post was requested as additional evidence, but its contents were not included in the supplied sources, so this profile does not describe specific 1.0 changes beyond what the homepage, README, and release metadata show.

What It Is

Pi is a terminal-based coding agent and agent harness. The README describes it as a minimal, extensible harness that can be customized with extensions, skills, prompt templates, and themes, which can be bundled as Pi packages and shared via npm or git. Pi ships with defaults but deliberately skips features such as sub-agents and plan mode, and the project suggests asking Pi to build them or installing a package.

It can be used interactively, in print or JSON mode for scripting, over RPC (JSON protocol over stdin/stdout), or embedded via a TypeScript SDK. The project points to OpenClaw as a real-world integration.

What the 1.0 release shows

The sources show that Pi 1.0 shipped together with Pi Durable. The monorepo now lists a @earendil-works/pi-durable package, described as a durable conversation, task, and document runtime. Other listed packages include a chord application-composition runtime, a telemetry contracts package, a unified multi-provider LLM API, an agent core runtime, and a terminal UI library with differential rendering.

The homepage also says MCP is now built in, with Codemode letting agents compose tool calls in a JavaScript sandbox; MCP extensions remain supported. This is a change from the earlier no-MCP stance, and the site links a post explaining why the team changed its mind. The README also describes a Nix run option, pinned-dependency installers, and supply-chain hardening practices.

Customization and extensibility

Extensions are TypeScript modules with access to tools, commands, keyboard shortcuts, events, and the TUI. The project lists example extensions for sub-agents, plan mode, permission gates, path protection, SSH execution, and sandboxing, and says more than 50 examples exist. After changing itself, Pi can be reloaded with /reload. Packages install with pi install from npm or git.

Models, sessions, and context

The homepage lists 15+ providers, including Anthropic, OpenAI, Google, Azure, Bedrock, Mistral, Groq, xAI, OpenRouter, and Ollama, authenticated by API key or OAuth. Models can be switched mid-session, and custom providers can be added through models.json or extensions.

Sessions are stored as trees in a single file, navigable with /tree, exportable to HTML with /export, and shareable via a GitHub gist with /share. For context engineering, Pi loads AGENTS.md, supports SYSTEM.md overrides, auto-compaction, on-demand skills, and prompt templates. Enter sends a steering message while the agent works, and Alt+Enter queues a follow-up.

Setup and security tradeoffs

Installation is via a curl or PowerShell installer, or npm, pnpm, or bun. Pi requires Node.js 22.19 or newer, and the installers can install it if needed. The README states Pi has no built-in permission system and runs with the permissions of the launching user, so it recommends containerizing or sandboxing it using Gondolin, Docker, or OpenShell. The README also notes that issues and PRs from new contributors are auto-closed by default.