A Piece of Pi: Embedding The OpenClaw Coding Agent In Your Product

Luebken connects the Unix idea of programs that do one thing well to building products for coding agents. He points to Claude’s finance tools as an example: a skill combines a CLI, Pandas, OpenPyXL, and LibreOffice tools to work with spreadsheets. He says there are no established patterns yet, so teams should experiment and keep systems accessible to agents.

Luebken describes Pi as an open-source, minimal starting point and explains that an agent runs tools in a loop using goals and context. His TypeScript CRM lead qualifier is a small terminal app that scores leads through tool calls. Hooks can check an action before a tool runs, while events let the application track tool results.

A coding agent adds a runtime and shell to the same tool-running loop. Luebken describes Pi extensions that connect session events and UI interactions: his CRM example adds a `/pipeline` command and lets users select contacts. He also shows a web UI built around the same extension mechanism, while noting that Pi’s framework is still geared toward coding.

Luebken explains that OpenClaw uses Pi’s core packages for embedded agents and sessions, while adding its own support for channels, provider orchestration, sub-agents, and gateways. In a client sales process, incoming proposal emails are routed to customer-specific agents with role and customer instructions, reusable sessions, and CLI access to CRM and ERP data. The system associates work with a case and produces an editable draft email for the user.