Agent · Pi
Pi: the agent that ships with the box.
Built by Mario Zechner. Wired into openclawOS by default. A serious tool-using LLM agent with memory, vision, and sub-agent delegation.
What Pi can do
- Tool use: shell, file system, web browser, code editor, calendar, vector memory, custom MCP servers.
- Vision: images sent through any channel get parsed — diagrams, screenshots, receipts.
- Sessions: per-sender, persistent, branchable. Switch devices, keep context.
- Compaction: when context fills up, Pi summarises older turns and continues.
- Sub-agents: delegate to specialised loops — researcher, writer, reviewer.
- Canvas UI: Pi can render structured UI back to iOS/Android nodes.
Recommended model
Pi is tuned for Claude Opus 4.6 (1M context) as the primary loop, with Sonnet for sub-agents. GPT-4o and Gemini 1.5 Pro work well too. For local-only, Llama 3.1 70B or Qwen 2.5 Coder 32B handle most coding tasks but tool-use reliability drops.
Where Pi lives
Pi is upstreamed from pi-mono. openclawOS embeds the same code — Pi runs in-process inside the Gateway with isolated per-binding state. You can fork Pi, patch a skill, and load your fork by pointing a config flag at the path.
Custom skills
A skill is a TypeScript function exposed as a tool. Drop one in ~/.openclaw/skills/
and Pi picks it up. Common community skills include: weather, smart-home, journaling,
calendar-ops, ticketing.
Welcome to the lobster tank.
Star the repo, join Discord, ship your first agent before lunch.