OpenClaw Architecture Overview

源码版本v2026.6.11

In one sentence

OpenClaw is a resident gateway + multi-channel ingress + agent main loop personal AI assistant. The gateway receives a message → feeds it to the agent → the agent repeatedly calls models and tools → the result is delivered back to each channel through the gateway. Every capability (tools / skills / plugins / providers / MCP / channels) is a pluggable registry.

Layers

Each layer in one line

• Startup and entry: openclaw.mjs performs a Node version check and forwards to src/entry.ts, which then dispatches subcommands via src/cli/run-main.ts.
• Gateway layer: startGatewayServer is the resident process. It accepts requests through an RPC method table and broadcasts agent events to WebSocket clients.
• Agent main loop: embedded-agent-runner's while(true) drives each conversation turn; a single attempt completes one LLM call and the tool_use / tool_result pairing.
• Capabilities: tools are a Map; skills are Markdown instructions that the model reads from SKILL.md; plugins are discovered by directory convention; providers are split by vendor; MCP is a bidirectional bridge.
• Channels: 22 channels (WhatsApp / Telegram / Slack ...) plug in via registry + loader + plugin, scheduled by the gateway.
• Config system: openclaw.json is the single source of truth; a Zod schema validates it at runtime.
• Memory and context: the Context Engine abstracts the compaction contract; memory files live at the workspace root.
• Scheduling and extensions: Cron runs jobs in an isolated agent session; Tasks persist via SQLite; ACP bridges IDEs.
• Deployment: the gateway is a resident process. The daemon wraps it as a system service, or you run it in Docker / Fly.

Why this layering

Why split it this way? OpenClaw fully decouples "where messages come from" (channels), "how they are processed" (agent loop), "which capabilities are used" (tools / skills / plugins / providers), "what persists" (config / memory / scheduling), and "how it runs" (deployment). Swapping a channel doesn't touch agent logic; swapping a model doesn't touch tools; adding a plugin doesn't require core changes. The gateway is the only hub — every cross-layer coordination goes through it.

Common misreadings

• "OpenClaw is a chat bot" — not quite. The bot is just one face of the channel layer; the gateway is the product core and the agent loop is the brain.
• "Tools and skills are the same thing" — they aren't. Tools are callable functions; skills are Markdown instructions that the model decides when to read and what to do with them.
• "MCP is only a client" — OpenClaw is both an MCP server (exposing its own capabilities externally) and a client (consuming external MCP servers).

Suggested reading order

Start with Startup and entry, then Gateway core and Agent main loop, then walk down through Capabilities → Channels → Config system → Memory → Scheduling → Deployment.

Official references: OpenClaw site · Official docs · DeepWiki.