Matthias Luebken, founder of Tavon.ai, presented a technical session at AI Engineer Europe in London on embedding OpenClaw’s Pi coding agent into third-party products. The talk, titled “A Piece of Pi: Embedding The OpenClaw Coding Agent In Your Product,” walked through Pi’s four runtime modes, OpenClaw’s plugin hook architecture, and a working CRM lead qualifier built on the framework.
Pi’s Four Runtime Modes
Pi, the minimalist coding agent created by Mario Zechner and now part of Earendil, ships with four distinct runtime modes according to its official documentation: interactive, print/JSON, RPC for process integration, and an SDK for embedding directly into applications. OpenClaw itself is the reference SDK integration, according to Pi’s documentation.
Luebken focused on the SDK and RPC modes as the entry points for product teams. Rather than running Pi as a standalone terminal tool, developers can instantiate agent sessions programmatically, inject custom tools, and control the full lifecycle from their own application code. The OpenClaw docs show the integration using @earendil-works/pi-coding-agent, with session creation, model registry, and tool injection handled through a structured API.
The Plugin Hook System
The core of Luebken’s presentation, according to StartupHub.ai, was OpenClaw’s plugin hook system. Designed for multi-channel, multi-agent platforms, the hooks cover multi-channel routing, model provider orchestration, sub-agent management, gateway lifecycle, session lifecycle, message persistence, observability, and production agent wrapping.
This modular architecture means a product team embedding Pi does not need to rebuild agent infrastructure from scratch. The hook system provides standardized extension points for each layer of the stack, from how messages are routed across channels to how agent sessions are monitored in production.
CRM Lead Qualifier Demo
Luebken demonstrated a “CRM Lead Qualifier” agent that scores and prioritizes sales leads by interacting with CRM data. The demo showed the agent searching contacts, scoring leads against configurable criteria, updating contact records, and logging interactions, all through Pi’s tool system. The underlying mechanism combined CLI-based data access with the agent’s four built-in tools: read, write, edit, and bash.
The webconsulting.at conference recap noted Luebken’s session as one of the practical application tracks, positioned alongside presentations from Mario Zechner on self-modifying software and Armin Ronacher on agent-legible architectures.
Pi’s Ecosystem by the Numbers
As of early May, Pi’s ecosystem has grown substantially. The Implicator reported that the badlogic/pi-mono repository held 45,041 GitHub stars under an MIT license, with the npm package at version 0.73.0 and 2,143 third-party packages in the catalog. Pi supports 15+ model providers, including Anthropic, OpenAI, Google, Mistral, Groq, xAI, and OpenRouter, with mid-session model switching.
The extension surface goes beyond tools. Pi packages can register commands, shortcuts, flags, custom renderers, widgets, dialogs, and event hooks, according to pi.dev. This is the extensibility layer Luebken argued makes Pi suitable as an embedded component rather than a standalone product.
Agent as Component, Not Product
Luebken’s closing message was direct: “Go Tinker.” The session framed Pi not as a competitor to Claude Code or Codex, but as a composable building block that product teams can shape to their domain. Where Claude Code and Codex ship editor, IDE, and cloud surfaces as default features, Pi treats each of those as something a user or package author adds. For teams building agent-powered products, that distinction determines whether they are integrating a tool or adopting a platform.