OpenClaw

A self-hosted AI agent platform — "an operating system for AI agents." Built around a central Gateway (Node.js 22+) using a hub-and-spoke architecture. Wraps the pi-agent-core runtime and supports multiple messaging platforms (WhatsApp, Telegram, Discord, Slack, iMessage) via channel adapters.

Capabilities

RPC surface: agent, agent.wait, and other gateway commands.
Session management: serialized per-session runs via session lanes + global queue.
Hook system: both internal gateway hooks and plugin hooks at fine-grained lifecycle points.
Streaming: bridges pi-agent-core events to structured assistant, tool, and lifecycle event streams.
Skills: loaded as a snapshot and injected into env + system prompt.
Multi-platform channels: WhatsApp, Telegram, Discord, Slack, iMessage via channel adapters.
Canvas / A2UI: agent-driven visual workspace with declarative UI component rendering.
Memory: hybrid search (vector + keyword) over long-term memory injected at context assembly.

Architecture

Channel Adapters (WhatsApp, Telegram, Discord…) ─┐
                                                  ├─→ Gateway (127.0.0.1:18789) ─→ Agent Runtime
Control Interfaces (Web UI, CLI, macOS app…)     ┘

Six-phase execution: ingestion → access control & routing → context assembly → model invocation → tool execution → response delivery.

Key concepts

OpenClaw Gateway — hub-and-spoke design, channel adapters, six-phase flow
Agent Loop (OpenClaw) — the core execution path (phases 3–5)
Context Assembly — AGENTS.md / SOUL.md / TOOLS.md + hybrid memory search
Canvas & A2UI — agent-driven visual workspace
Security Model — layered defense: loopback, auth, device pairing, Docker sandboxing
Deployment Options — local, macOS LaunchAgent, Linux/VPS, Fly.io
Command Queue — queue modes (collect/steer/followup) and lane system
Streaming — chunking and block reply behavior
Compaction — long-context summarization
OpenClaw Hooks — internal gateway hooks