Pi: A Tiny, Self-Extending Agent That Builds Its Own Tools

Pi is a minimal coding agent powering OpenClaw, built on the idea that LLMs should write and run code while the agent stays small. It avoids heavy built-in integrations like MCP, instead using extensions, skills, CLIs, and a robust TUI with hot reloading, session state, and branching. The author uses Pi to generate and maintain bespoke capabilities, arguing this self-extending approach is the future.

Key Points

• Pi embraces a tiny, reliable core (four tools, ultra-short prompt) and a powerful self-extension model where the agent writes and hot-reloads its own capabilities.
• By design, Pi does not include MCP; instead it prefers CLIs/skills and can use tools like mcporter if needed, keeping most functionality outside the LLM tool context.
• Sessions are tree-structured with branching and rewinding, persist custom state across extensions, and can mix providers—minimizing provider-specific lock-in.
• A rich TUI system enables practical, interactive workflows (/answer, /todos, /review, /control, /files), making agent interactions observable and ergonomic.
• The author relies on Pi to author bespoke skills (e.g., CDP-based browsing, commit/changelog rules, uv routing), exemplifying software that builds more software and informing OpenClaw’s design.