# Pi Coding Agent

A minimal terminal coding agent by Mario Zechner. Pi strips away the feature bloat common in AI coding tools — no built-in MCP, no sub-agents, no permission popups, no plan mode, no to-dos. Instead, it gives you a small, fast core and expects you to extend it yourself via TypeScript modules and npm packages.

The extend-it-yourself philosophy produced a real community ecosystem, and two "batteries-included Pi" distributions now sit on top of it: [[oh-my-pi]], a maximalist *fork* that bakes tools into the binary, and [[lazypi|LazyPi]], a one-command *installer* that configures unmodified Pi with 60+ community skills, 67 themes, MCP adapter, sub-agents, and persistent memory. The two are the fork-vs-installer answers to the same "vanilla Pi ships empty" gap.

**Install:** `npm install -g @mariozechner/pi-coding-agent`

**Providers:** supports 15+ — Anthropic, OpenAI, Google, Azure, Bedrock, Mistral, Groq, Cerebras, xAI, Hugging Face, Kimi, MiniMax, OpenRouter, Ollama, and others.

**Four modes of operation:**
- Interactive TUI for normal use
- Print/JSON mode for scripting
- RPC (JSON over stdin/stdout) for tool integration
- SDK for embedding in other applications

**Session management** uses a tree structure — you can branch conversation history, share sessions, and navigate back to earlier points. Auto-compaction and dynamic context injection handle long sessions.

**Extensibility** is the main selling point. Skills, prompt templates, themes, and tools are all TypeScript modules you can bundle and distribute via npm or git. The philosophy is that features like MCP support or sub-agents should be community packages, not baked into the core.

**Repo:** https://github.com/badlogic/pi-mono/tree/main/packages/coding-agent — TypeScript, MIT license.