How OpenClaw Is Built: Architecture Explained

High-Level Overview

OpenClaw follows a client-server architecture with three main components:

CLI (Command Line Interface)

The CLI is your primary interface for managing OpenClaw. Install via npm, pnpm, or bun:

npm install -g openclaw@latest

Common commands:

• • openclaw gateway start — Start the gateway
• • openclaw configure — Configure settings
• • openclaw status — Check status
• • openclaw chat — Start a chat session
• • openclaw security audit — Run security checks

Gateway

The gateway is the core service that powers OpenClaw. It handles:

• •Message routing — Routes messages between channels and agents
• •Agent orchestration — Manages agent state and memory
• •Tool execution — Runs tools on behalf of agents
• •Channel connections — Connects to Telegram, Discord, WhatsApp, etc.
• •Authentication — Validates tokens and permissions

The gateway runs as a local service (launchd on macOS, systemd on Linux) and exposes a WebSocket API on port 18789 by default.

Control UI / Dashboard

The web-based dashboard for managing OpenClaw. Access at:

http://127.0.0.1:18789

Features:

• • Agent configuration and monitoring
• • Channel management
• • Real-time activity logs
• • Configuration editor
• • Security settings

Models & Providers

OpenClaw connects to LLM providers through a unified API layer:

Configure providers in your openclaw.json config file.

Skills & Tools Layer

Skills extend OpenClaw's capabilities. They define what agents can do:

• • File operations — Read, write, search files
• • Web search — Search the web for information
• • Browser control — Browse websites, fill forms
• • Code execution — Run code in sandboxes
• • Custom skills — Build your own tools

Integrations / Channels

OpenClaw connects to messaging platforms through channel adapters:

Config, Memory & State

OpenClaw persists data in ~/.openclaw/:

• • openclaw.json — Main configuration
• • credentials/ — Channel API keys
• • agents/ — Agent memory and sessions
• • workspace/ — Agent working files
• • skills/ — Custom skills

Local vs Remote Deployment

Security

OpenClaw is designed as a personal assistant — not a multi-tenant system. Key security points:

• • Gateway binds to loopback by default
• • Token-based authentication
• • Channel allowlists restrict who can message
• • Tool permissions control what agents can do
• • Elevated mode for sensitive operations