OpenClaw Documentation
by OpenClaw (Peter Steinberger and contributors)
The reference for the self-hosted gateway pattern — chat app in, agentic loop and skills out, running on your own box.
Overview
OpenClaw is a free, MIT-licensed autonomous agent that uses messaging platforms as its user interface, created by Austrian programmer Peter Steinberger. It launched in November 2025 as Warelay, was renamed Moltbot on 27 January 2026 after a trademark dispute with Anthropic, then became OpenClaw three days later; by 2 March 2026 it had roughly 247,000 GitHub stars and 47,700 forks, making it one of the fastest-growing open-source projects of the year. It runs as a single self-hosted process on your own hardware and calls out to external providers such as Claude, DeepSeek or GPT. The documentation is organised into ten hubs: Get Started (overview, showcase, setup), Install (installation paths, containers, hosting), Channels (messaging integrations, pairing, routing), Agents (architecture, sessions, context, memory), Capabilities (tools, skills, automation, webhooks), ClawHub (plugin marketplace and publishing), Models (providers, configuration, failover), Platforms (macOS, Windows, iOS, Android, web), Gateway and Ops (configuration, security, diagnostics) and Reference (CLI, schemas, RPC, release notes). Quick start is a single shell command — curl -fsSL https://openclaw.ai/install.sh | bash on macOS/Linux/WSL2, or an iwr one-liner on Windows PowerShell — with Node 26 (or 24.16+/26.1+) and one provider API key as the only requirements. Skills follow the now-common directory-of-metadata-plus-instructions convention. Read the security pages before exposing it: documented criticisms include broad permission requirements, prompt-injection exposure and unvetted third-party skills, and Cisco researchers found a community skill performing data exfiltration and prompt injection without user awareness.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Documentation
- Cost
- Free
- Duration
- ~2-4 hours to install the gateway, pair a channel and ship a first skill; ongoing as a reference
- Provider
- OpenClaw (Peter Steinberger and contributors)
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓How to install and run a self-hosted agent gateway as a single local process
- ✓How to pair and route across ten messaging channels including Signal, Slack and WhatsApp
- ✓How the agent architecture separates sessions, context assembly and persistent long-term memory
- ✓How to write, install and publish skills and plugins through the ClawHub marketplace
- ✓How to configure multiple model providers with failover so one outage does not stop the agent
- ✓How to operate the gateway: configuration files, diagnostics, security hardening and the CLI
- ✓How webhooks and automation triggers let the agent act without a human message first
- ✓Why unvetted third-party skills are the main attack surface, and what permissions to withhold
Highlights
- •Documents a genuinely different agent shape — messaging app as the interface, not a terminal or an IDE
- •Complete operational surface is covered, not just the happy path: gateway config, diagnostics, RPC, schemas and release notes all have reference pages
- •Multi-provider model configuration with explicit failover is documented as a first-class concern rather than an afterthought
- •The security story is documented rather than hidden, which matters because researchers have demonstrated real exfiltration through community skills
- •Runs entirely on your own hardware with an MIT licence, so the whole agent loop is inspectable
Who It’s For
Best For
- ✓Engineers who want a persistent personal agent reachable from a phone rather than a laptop terminal
- ✓Teams designing multi-channel agent gateways and needing a working reference architecture
- ✓Skill and plugin authors targeting a large existing agent ecosystem
- ✓Security engineers studying prompt-injection and permission design in a widely deployed agent
Prerequisites
- •Node 26 (or 24.16+/26.1+) and comfort self-hosting a long-running process
- •An API key from an LLM provider such as Anthropic, OpenAI or DeepSeek
- •Working understanding of prompt injection risk before granting an agent tool or file access
FAQ
What is OpenClaw Documentation?
Documentation for OpenClaw, the self-hosted gateway that puts an LLM agent behind ordinary messaging apps — Discord, Slack, Signal, Telegram, WhatsApp, iMessage, Matrix, Microsoft Teams, Google Chat and Zalo. It is aimed at engineers who want to run a persistent, memory-backed agent on their own hardware, and it covers channel routing, session and memory architecture, skills, plugins, model failover and the operational security you have to get right.
Is OpenClaw Documentation free?
OpenClaw Documentation is free to access.
What level is OpenClaw Documentation for?
OpenClaw Documentation is aimed at a intermediate audience. Recommended background: Node 26 (or 24.16+/26.1+) and comfort self-hosting a long-running process, An API key from an LLM provider such as Anthropic, OpenAI or DeepSeek, Working understanding of prompt injection risk before granting an agent tool or file access.
How long does OpenClaw Documentation take?
Expect roughly ~2-4 hours to install the gateway, pair a channel and ship a first skill; ongoing as a reference. Most learners work through it at their own pace.
What will I learn from OpenClaw Documentation?
You'll learn: How to install and run a self-hosted agent gateway as a single local process; How to pair and route across ten messaging channels including Signal, Slack and WhatsApp; How the agent architecture separates sessions, context assembly and persistent long-term memory; How to write, install and publish skills and plugins through the ClawHub marketplace; How to configure multiple model providers with failover so one outage does not stop the agent; How to operate the gateway: configuration files, diagnostics, security hardening and the CLI; How webhooks and automation triggers let the agent act without a human message first; Why unvetted third-party skills are the main attack surface, and what permissions to withhold.
Topics
Sources
This page was written from 2 sources, 1 on domains other than docs.openclaw.ai.