What Is OpenClaw? The Open-Source Personal AI Assistant You Run On Your Own Devices
OpenClaw is an open-source personal AI assistant that runs locally on your hardware and answers you through WhatsApp, Telegram, Signal, or any chat app you already use. It reached 247k GitHub stars in weeks. Here is what it is, how it works, and how to get started.
In November 2025, Peter Steinberger — the developer behind the PDF SDK used in Dropbox, Slack, and Box — pushed a side project to GitHub called Warelay. He was burned out on SaaS and wanted a personal AI that ran on his own machines, answered him through the chat apps he already used, and did not hand his data to a third party.
By March 2026, it had 247,000 GitHub stars and 47,700 forks. Renamed twice (first to Moltbot, then to OpenClaw after a trademark dispute), it became one of the fastest-growing open-source projects ever recorded. Steinberger joined OpenAI in February 2026 to work on bringing agents to everyone; OpenClaw moved to an independent foundation and stayed open.
This is the complete guide to what OpenClaw is, how it works, and how to set it up.
Update (July 9, 2026 — Foundation): OpenClaw is now a 501(c)(3) non-profit with full-time staff and major partners. OpenClaw Foundation guide →
OpenClaw is a local-first AI gateway that runs on your machine, connects to any messaging channel you already use, and extends what it can do through installable claws.
You bring the AI model (Claude, GPT-5.5, DeepSeek, a local Ollama model — your choice). OpenClaw handles the routing, memory, sessions, and tool execution. Your messages go in through Telegram or Signal or WhatsApp; answers come back the same way.
Why it took off
Two problems that most AI products do not solve:
The interface problem. Most AI assistants force you into a web interface or a dedicated app. OpenClaw answers you on the channels you are already in. You message your Telegram bot at 2am and get a real answer. No new app to open.
The data problem. Every message you send to a hosted AI is training data for someone else's model or is stored on someone else's server. OpenClaw runs on your machine. The conversation history stays local. The model call goes to your API key, not through a middleman.
Those two properties — answer me where I already am, don't touch my data — turned out to be exactly what a large slice of developers and power users wanted but could not get anywhere else.
Channels — the messaging surfaces where you send and receive messages. Telegram, Signal, WhatsApp, Discord, Slack, iMessage, and 20+ others. Each channel registers as an inbox. One OpenClaw instance can handle messages from all of them at once.
Gateway — the local control plane. Routes each incoming message to the right session and agent, manages tool execution, handles memory, and streams responses back to the channel. Runs as a Node.js process on your machine.
Skills (claws) — installable capability packs that extend what the agent can do. Each claw is a directory with a SKILL.md file: YAML frontmatter naming the skill, plus natural-language instructions that the model reads to understand how to use it. Claws can wrap web search, file access, API calls, CLI tools, or any workflow you can describe in markdown.
The model — Claude, GPT-5.5, DeepSeek, or a local Ollama model — reads your message, reads any relevant skill instructions, executes tool calls (web search, file read, shell command, API call), and returns a response. All of this runs locally except the model inference itself, which goes to your chosen provider's API.
What a claw actually is
A claw is a markdown file with instructions. Here is the anatomy of a minimal one:
markdown
---
name: github
description: Read and comment on GitHub pull requests, issues, and repositories.
tools: [web_fetch, exec]
---
When the user asks about a GitHub PR or issue, fetch the URL and read the content.
Use `gh` CLI for write actions — opening PRs, posting comments, approving reviews.
Always confirm before making write operations.
When you ask something like "summarize the open PRs on my repo", OpenClaw matches your request to the github skill's description, loads the instructions into context, and follows them. You do not have to invoke skills by name — the model figures out which skill applies based on your request.
Skills can be global (available to all agents) or workspace-scoped (only for a specific agent config). They can include supporting files — scripts, templates, reference docs — alongside the SKILL.md.
Supported models and providers
OpenClaw is model-agnostic. At the gateway config level, you point it at any provider:
Provider
Notes
Anthropic Claude
Claude 4.x / Fable 5 — best for complex reasoning and long context
OpenAI GPT-5.5 / GPT-5.5-Cyber
Strong general performance
DeepSeek
Cost-efficient alternative; popular in the community
Google Gemini
Long-context option
Ollama
Run models locally — no API cost, fully offline
Any OpenAI-compatible endpoint
Custom deployments, self-hosted models
You switch models per-agent or per-workspace — your personal assistant can run on Claude while a coding agent runs on DeepSeek for cost reasons.
Setting up OpenClaw
Requirements
Node 24 (recommended) or Node 22 LTS (22.19+)
An API key for your chosen AI model
5 minutes
Install and initialize
bash
npm install -g openclaw
# or
npx openclaw@latest onboard
The onboard command walks you step by step through:
Choosing a model provider and entering your API key
Setting up your first channel (Telegram is recommended for first-timers)
Configuring a workspace directory
Installing a starter set of claws
Set up Telegram (fastest path)
Telegram is the simplest channel to start with:
Message @BotFather on Telegram → /newbot → follow prompts → copy your bot token
In your OpenClaw config, set TELEGRAM_BOT_TOKEN=<your_token>
Run openclaw serve
Message your bot
Your messages go to your local OpenClaw process. The gateway handles the session and routes to your configured model.
Set up Signal, WhatsApp, or other channels
Each channel has its own transport configuration. Signal requires a Signal registration; WhatsApp needs a Baileys bridge or Meta Cloud API access. See the official docs at docs.openclaw.ai for channel-specific setup.
Installing claws
ClawHub (clawhub.ai) is the official skill registry with 2,857+ claws (post-security audit). Install via CLI:
bash
# Install into current workspace
openclaw skills install github
# Install globally (available to all agents)
openclaw skills install github --global
# Browse categories
openclaw skills search "web search"
openclaw skills search "memory"
The ClawHub UI lets you browse by category, sort by downloads and stars, and read changelogs. Every skill has a VirusTotal scan report — check it before installing community claws, especially for write-enabled skills.
Core built-in tools
Before installing any claws, OpenClaw ships with a base set of tools:
Tool
What it does
read
Read files from your filesystem
write
Write and edit files
exec
Run shell commands
web_search
Search the web (12 providers, auto-selects based on available API keys)
web_fetch
Fetch and read any URL
browser
Control a browser — click buttons, fill forms, take screenshots
memory
Store and recall facts across sessions
These tools are what claws can call. When you install a claw, it declares which tools it uses in its tools: frontmatter field — so you know upfront what filesystem or network access a skill requires.
Web search providers
OpenClaw supports 12 web search providers. If no provider is explicitly configured, it auto-detects based on which API keys you have set, checking in this priority order:
DuckDuckGo requires no API key and works out of the box. Brave and Perplexity are the most popular community choices for quality. Firecrawl and Exa are popular for deep research tasks.
Memory and sessions
OpenClaw maintains session state across messages within a conversation. The memory system stores facts, preferences, and context in local SQLite — not on any external server.
Claws like memory-wiki (one of the highest-rated on ClawHub) teach the agent patterns for proactively storing and recalling context: your preferences, project details, recurring facts. Agents with good memory setups adapt to your workflow over time in ways that stateless chat UIs cannot.
BirdClaw and OpenClaw: how they relate
BirdClaw is Peter Steinberger's separate project for local Twitter/X management — a local SQLite database for your tweets, DMs, likes, and bookmarks with AI digests and a clean web UI. It ships with its own SKILL.md in its .agents/skills/birdclaw directory, which means BirdClaw-aware tasks can be wired into OpenClaw as a claw.
The two projects share the same local-first philosophy and the same author. BirdClaw handles the data layer for your Twitter history; an OpenClaw agent with the BirdClaw claw can query it, triage mentions, and generate digests through your preferred messaging channel.
In January 2026, researchers at Koi Security uncovered ClawHavoc — a coordinated attack campaign that planted 341 malicious skills on ClawHub using typo-squatting. ClawHub removed 2,419 suspicious skills, partnered with VirusTotal for automatic scanning, and the registry shrank from 5,705 to ~3,286 clean skills.
The practical guidance:
Install skills from verified publishers (openclaw org, well-known authors with high star counts)
Check the VirusTotal report on any skill before installing
Pay attention to which tools a skill declares — a skill claiming to need exec for what should be a read-only task is a red flag
Prefer workspace-scoped installs over --global for skills from unknown publishers
OpenClaw vs alternatives
OpenClaw
Claude Desktop
ChatGPT Desktop
n8n
Where you interact
Any chat app you already use
Claude Desktop app
ChatGPT app
Web UI / webhooks
Data stays local
✅
Partial
❌
Self-hosted option
Model choice
Any
Claude only
OpenAI only
Any
Skill/plugin ecosystem
5,400+ claws
MCP servers
GPTs
400+ nodes
Self-hosted
✅
❌
❌
✅
Cost
API cost only
Claude subscription
OpenAI subscription
Free/paid
The strongest OpenClaw argument is the channel flexibility and the bring-your-own-model approach. If you live in Telegram and want Claude to answer you there — with access to your files, web search, and custom skills — OpenClaw is the clearest path to that.
What it can do out of the box
Without installing any claws, a fresh OpenClaw with a good model can:
Answer questions using web search (built-in)
Read and write local files
Run shell commands with confirmation
Fetch and summarize web pages
Remember context across a conversation session
With popular claws installed:
Manage GitHub PRs and issues
Summarize and search your Obsidian vault
Triage Twitter/X with BirdClaw
Send Slack messages
Read and draft emails (Gmail / Outlook claws)
Control a browser for form-filling and web automation
Search your documents with Notion or Confluence claws
Getting started now
bash
# Install
npm install -g openclaw
# Start onboarding
openclaw onboard
# After setup: start the gateway
openclaw serve
# Install your first claw
openclaw skills install github --global
openclaw skills install web-search --global
