Pi Coding Agent

Pi Coding Agent is an MIT-licensed terminal coding harness maintained in the Earendil Works Pi monorepo. It starts with a deliberately small core: an interactive terminal UI and four default tools for reading, writing, editing, and running shell commands. Pi then exposes extensions, skills, prompt templates, themes, and installable Pi packages so users can add their own commands, tools, UI, events, policies, or integrations without modifying the core. It also runs in print, JSON-event, JSON-RPC, and SDK modes.

Pi is installed from the @earendil-works/pi-coding-agent npm package or, on Linux and macOS, through the official shell installer. The docs include Windows and Termux-on-Android guidance. Authentication uses supported provider subscriptions, API keys, or configured local/custom endpoints; model charges and subscription terms therefore come from the selected provider rather than a separate Pi plan. The project explicitly notes that features such as subagents and plan mode are examples or packages, not always-on core features.

Top features

• Minimal terminal agent: repository-aware read, write, edit, and shell tools are available in interactive coding sessions.
• Tree-structured history: saved JSONL sessions support resume, branching, navigation, compaction, export, and sharing workflows.
• Runtime model switching: users can authenticate to supported providers and change models or thinking settings during a session.
• Context and skills: AGENTS.md or CLAUDE.md files, skill packages, prompt templates, and themes customize project behavior.
• TypeScript extensions: add commands, tools, key bindings, UI components, event handlers, policy gates, or custom compaction.
• Programmatic operation: print mode, JSON event streams, an SDK, and stdin/stdout JSON-RPC support scripts and integrations.
• Optional isolation: official guidance covers Gondolin, Docker, and OpenShell patterns when stronger boundaries are needed.

Use cases

• Edit and test a repository from a terminal with a small, inspectable set of default tools.
• Create a read-only review profile by limiting the tools available to the model.
• Package team prompts, skills, themes, and extensions for reuse across projects.
• Embed an agent session in a Node application or control it from another language through JSON-RPC.
• Add custom approval gates, protected paths, SSH execution, or subagent behavior as explicit extensions.