What Is OpenClaw? The Complete Guide (2026)
The open-source AI agent with 187K GitHub stars that turns your messaging apps into command centers. Here's everything you need to know before installing it.
OpenClaw is a free, open-source AI agent that runs on your own computer and connects to messaging platforms like WhatsApp, Telegram, Discord, Slack, and more than 15 other channels. You send it a message -- in plain language -- and it executes tasks using AI. Think of it as a personal assistant that lives inside the apps you already use, powered by whichever AI model you choose: Claude, GPT, Gemini, Grok, or even a local model running offline.
With 187,000+ GitHub stars, 20,000+ forks, and a Discord community of 60,000 members, OpenClaw has become the most popular open-source AI agent project in the world. It's also one of the most controversial, thanks to a major security vulnerability, an Anthropic trademark dispute, and an AI social network that grew to 1.5 million agents.
The Rebrand Story: Clawdbot to Moltbot to OpenClaw
OpenClaw wasn't always called OpenClaw. Understanding the name changes helps you follow discussions online, since many people still use the old names.
Clawdbot (early 2025)
The project launched as Clawdbot -- a portmanteau of "Claude" and "bot." It was built specifically around Anthropic's Claude API and wore the reference openly in its name.
Moltbot (mid 2025)
Anthropic raised trademark concerns about the name's similarity to "Claude." The project rebranded to Moltbot, a reference to molting -- shedding an old shell. The name stuck for several months and still appears in older tutorials and forum posts.
OpenClaw (late 2025 -- present)
As the project added support for GPT, Gemini, Grok, and local models via Ollama, the Claude-specific branding no longer fit. The final rebrand to OpenClaw emphasized two things: it's open source and model-agnostic. The "Claw" preserved continuity with the original name while dropping the direct Claude reference.
Who Created OpenClaw?
Peter Steinberger built OpenClaw as a solo project. He's an Austrian developer based in Vienna, best known as the founder of PSPDFKit, a widely-used PDF framework for mobile and web applications.
Steinberger started OpenClaw as a personal experiment to connect Claude to his messaging apps. The project's rapid growth -- from a weekend side project to 187K GitHub stars -- was driven almost entirely by the open-source community contributing skills, integrations, and platform adapters.
OpenClaw is released under the MIT license, meaning anyone can use, modify, and distribute it freely, including for commercial purposes.
How OpenClaw Works: The Gateway Architecture
OpenClaw's design is straightforward once you understand the three layers.
1. The messaging layer (channels)
OpenClaw connects to messaging platforms through channel adapters. Each adapter translates platform-specific protocols into a common format. When you send a WhatsApp message, the WhatsApp adapter receives it and passes it to the core in a standardized way.
2. The core (agent loop)
The core is where the AI thinking happens. It follows an agent loop:
- Receive your message
- Check persistent memory for context about you and the conversation
- Decide which skills (if any) to invoke
- Send a prompt to the AI model (Claude, GPT, etc.)
- Execute any actions the model requests (web searches, file operations, API calls)
- Return the result to you through the same messaging channel
This is a messaging-first design. Unlike tools that require you to open a separate app or website, OpenClaw meets you where you already are -- inside your chat apps.
3. The skills layer
Skills are plugins that extend what OpenClaw can do. The community has built 5,705 skills (and counting) on ClawHub, the project's skill marketplace. Skills handle everything from sending emails and managing calendars to generating images and scraping websites.
Key Features of OpenClaw
Multi-channel messaging
Connect to 15+ messaging platforms from a single installation. Send a message on WhatsApp, continue the conversation on Telegram, and check results on Discord. OpenClaw maintains context across channels.
5,700+ community skills
ClawHub hosts 5,705 skills that extend OpenClaw's capabilities. Categories include productivity, automation, data analysis, content creation, communication, developer tools, and more. Installing a skill is a single command.
Persistent memory
OpenClaw remembers your preferences, past conversations, and context between sessions. Unlike chatbots that reset with every conversation, OpenClaw builds a growing understanding of your needs over time.
Browser automation
Built-in browser control lets OpenClaw navigate websites, fill forms, extract data, and take screenshots. This runs headlessly (in the background) or visually so you can watch what it's doing.
Model-agnostic
OpenClaw works with multiple AI providers. You choose which model powers your agent and can switch between them depending on the task. More on supported models below.
Cron jobs and heartbeat
Schedule recurring tasks with cron-style triggers. Set OpenClaw to check your email summary every morning, generate a daily report at 5 PM, or monitor a website for changes every hour. The heartbeat system keeps background tasks alive even when you're not actively messaging.
Canvas
Canvas is OpenClaw's visual workspace -- a web-based interface for tasks that don't fit neatly into a chat message. Use it for viewing generated documents, editing structured data, or working with images and charts that OpenClaw produces.
50+ integrations
Beyond messaging platforms, OpenClaw connects to 50+ external services: Google Workspace, Notion, GitHub, Trello, Airtable, Zapier, and others. These integrations let OpenClaw read from and write to the tools your business already uses.
Supported Messaging Platforms
OpenClaw's core selling point is that it works inside your existing messaging apps. Here's the full list of supported channels:
| Platform | Status | Notes |
|---|---|---|
| Stable | Personal and Business API | |
| Telegram | Stable | Bots and user accounts |
| Discord | Stable | Server bots, DMs, threads |
| Slack | Stable | Workspace apps |
| Signal | Stable | Via signal-cli bridge |
| iMessage | Beta | macOS only, uses AppleScript bridge |
| Microsoft Teams | Stable | Requires Azure bot registration |
| Stable | IMAP/SMTP, any provider | |
| Matrix | Stable | End-to-end encrypted |
| Messenger | Beta | Facebook/Meta pages |
| LINE | Stable | Popular in Japan/SE Asia |
| Community | Unofficial adapter | |
| SMS | Stable | Via Twilio |
| Web Chat | Stable | Embeddable widget |
| Webhook | Stable | Custom HTTP integrations |
Most users start with WhatsApp or Telegram since setup is the simplest. You can add more channels later without reconfiguring anything. See our WhatsApp & Telegram setup guide for a step-by-step walkthrough.
AI Models Supported by OpenClaw
OpenClaw is model-agnostic. You bring your own API key and choose which model handles your requests.
| Provider | Models | Best For |
|---|---|---|
| Anthropic | Claude Opus 4.6, Sonnet 4.5, Haiku 4.5 | Complex reasoning, long tasks, coding |
| OpenAI | GPT-4o, GPT-4.5, o3 | General purpose, vision, function calling |
| Gemini 2.0 Flash, Gemini 2.0 Pro | Speed, multimodal, large context | |
| xAI | Grok 3 | Real-time info, unfiltered responses |
| Local (Ollama) | Llama, Mistral, Phi, any GGUF | Privacy, offline use, zero API costs |
You can also set different models for different tasks. For example, use Claude Opus 4.6 for complex reasoning and Gemini Flash for quick lookups, keeping your costs down while maintaining quality where it matters.
How to Install OpenClaw: Quick Start
OpenClaw offers three installation methods. Pick the one that matches your comfort level.
Option 1: One-line install (recommended)
The fastest way to get started on Mac or Linux:
curl -fsSL https://openclaw.sh/install | bash
This downloads OpenClaw, installs dependencies, and walks you through connecting your first messaging channel.
Option 2: npm
If you already have Node.js installed:
npm install -g @openclaw/cli
openclaw init
Option 3: Docker
For isolated environments or servers:
docker run -d --name openclaw \
-v openclaw-data:/data \
-p 3000:3000 \
openclaw/openclaw:latest
After installation, you'll need to add your AI provider API key and connect at least one messaging channel. The setup wizard handles both steps.
OpenClaw Pricing: What It Actually Costs
This trips people up, so let's be precise. OpenClaw has three cost components, and only one is required.
1. The software: free
OpenClaw itself costs nothing. It's MIT-licensed open-source software. You download it, run it, and modify it without paying anyone.
2. AI model API costs: $5-150/month (required)
You need an API key from at least one AI provider. How much you spend depends on which model you use and how much you chat:
| Usage Level | Model | Estimated Monthly Cost |
|---|---|---|
| Light (10-20 messages/day) | Claude Haiku / Gemini Flash | $5-15 |
| Moderate (50-100 messages/day) | Claude Sonnet / GPT-4o | $30-60 |
| Heavy (200+ messages/day) | Claude Opus / GPT-4.5 | $80-150+ |
| Local models (Ollama) | Llama, Mistral | $0 (your hardware) |
3. Managed hosting: $6-149/month (optional)
If you don't want to run OpenClaw on your own computer, several managed hosting providers will run it for you. Plans range from $6/month for basic personal use to $149/month for business accounts with multiple channels, priority support, and guaranteed uptime.
Self-hosting is free beyond the electricity your computer uses. Most individuals self-host on a Raspberry Pi, an old laptop, or a $5/month cloud server.
Security: The Elephant in the Room
OpenClaw's biggest controversy has been security, and it's important to understand what happened and where things stand.
CVE-2026-25253: The wake-up call
In January 2026, security researchers disclosed CVE-2026-25253, a critical vulnerability in OpenClaw's skill system (Source: NVD/MITRE, CVE-2026-25253). Malicious skills could escape the sandbox and execute arbitrary code on the host machine. This meant that installing a compromised skill from ClawHub could give an attacker full access to your computer.
The vulnerability was patched within 48 hours, but the damage to trust was significant.
Malicious skills on ClawHub
Following the CVE disclosure, auditors found several skills on ClawHub that contained hidden data-exfiltration code. These skills appeared legitimate on the surface -- a weather plugin, a note-taking helper -- but quietly sent user data to external servers.
Enterprise bans
Multiple companies banned OpenClaw from corporate devices after the vulnerability disclosure (via Korean tech press and The Register). IT security teams flagged the combination of broad system access, community-contributed skills, and messaging platform integration as too high-risk for enterprise environments.
What's been done since
- VirusTotal partnership -- All skills on ClawHub are now scanned for malware and suspicious behavior before publishing
- Sandboxing improvements -- Skills now run in isolated containers with explicit permission grants
- Verified publisher badges -- ClawHub now distinguishes between verified and unverified skill authors
- Audit logging -- Every action a skill takes is logged and reviewable
The Moltbook Phenomenon: An AI Social Network
One of the most unexpected developments in the OpenClaw ecosystem is Moltbook (the name predates the rebrand) -- a social network where AI agents interact with each other.
Users give their OpenClaw agents public profiles, personalities, and the ability to post, comment, and interact with other agents. As of February 2026, Moltbook hosts approximately 1.5 million AI agents.
Some agents curate news feeds. Others write poetry. Some run entire fictional personas with backstories and opinions. It's equal parts fascinating and surreal -- a glimpse of what happens when AI agents are given social presence.
Moltbook is not required to use OpenClaw. It's a community-built layer on top of the core platform. But it's a significant part of the culture and ecosystem around the project.
OpenClaw vs Claude Code: Two Different Tools
This comparison comes up constantly, so let's be clear: OpenClaw and Claude Code solve different problems.
| Dimension | OpenClaw | Claude Code |
|---|---|---|
| Primary interface | Messaging apps (WhatsApp, Telegram, etc.) | Terminal / IDE |
| Primary purpose | Personal AI assistant across platforms | Code execution and file operations |
| Who makes it | Open-source community (Peter Steinberger) | Anthropic (official product) |
| AI models | Any (Claude, GPT, Gemini, Grok, local) | Claude only (Anthropic models) |
| Best for | Messaging-based workflows, multi-channel | Building, coding, file manipulation |
| Extensibility | 5,700+ ClawHub skills | MCP servers, custom skills, hooks |
| Runs on | Self-hosted or managed cloud | Your local machine |
| License | MIT (fully open source) | Proprietary (Anthropic) |
The tools are complementary, not competing. Many power users run both: Claude Code for building projects and working with files, and OpenClaw for staying connected to AI through their messaging apps throughout the day.
If you work at a desk and need an AI that manipulates code and files, Claude Code is the better fit. If you want an AI assistant you can message from your phone while walking to lunch, OpenClaw is what you want.
Who Should Use OpenClaw
OpenClaw is a strong fit if you:
- Live in messaging apps -- you already spend hours in WhatsApp, Telegram, or Slack and want AI where you are
- Want a personal AI assistant that remembers your preferences and builds context over time
- Need multi-platform reach -- same AI, accessible from your phone, desktop, and any messaging app
- Value model flexibility -- you want to switch between Claude, GPT, Gemini, or local models without changing tools
- Like automating recurring tasks -- daily summaries, scheduled reports, website monitoring
- Enjoy tinkering -- OpenClaw rewards curiosity with a deep skill ecosystem and active community
Who should skip it
- Enterprise/compliance-heavy environments -- the security history makes it a tough sell to IT departments
- People who want zero setup -- even the one-line installer requires some configuration
- Users who need guaranteed uptime -- self-hosted means you're responsible for keeping it running
- Anyone uncomfortable with open-source risk -- community skills carry inherent trust issues despite recent improvements
Getting Started with OpenClaw: Your Checklist
Ready to try it? Here's the step-by-step path from zero to working agent.
- Choose your AI provider. Sign up for an API key from Anthropic, OpenAI, or Google AI. Add $10-20 in credits to start.
- Install OpenClaw. Run
curl -fsSL https://openclaw.sh/install | bashon Mac/Linux. See our installation guide for Windows and Docker options. - Connect a messaging channel. Start with Telegram (easiest) or WhatsApp. The setup wizard walks you through creating a bot and linking it.
- Send your first message. Open the connected app and send "Hello" to your bot. You should get an AI-powered response within seconds.
- Install your first skill. Try something simple like a weather or reminder skill. Run
openclaw skill install weatherfrom the CLI. - Set up persistent memory. Tell your agent your name, preferences, and common tasks. It will remember these across sessions.
- Explore ClawHub. Browse the 5,700+ available skills -- but stick to verified publishers.
- Join the community. The OpenClaw Discord (60K members) is the best place for troubleshooting and discovering new use cases.
OpenClaw by the Numbers (February 2026)
| Metric | Number |
|---|---|
| GitHub stars | 187,000+ |
| GitHub forks | 20,000+ |
| Discord members | 60,000+ |
| ClawHub skills | 5,705 |
| Integrations | 50+ |
| Messaging channels | 15+ |
| Website visits (peak week) | 2,000,000+ |
| Moltbook AI agents | 1,500,000 |
| License | MIT (free and open source) |
| Creator | Peter Steinberger (PSPDFKit) |
The Bottom Line
OpenClaw is the most ambitious open-source AI agent project in existence. It puts a powerful, customizable AI assistant inside the messaging apps you already use, and it does so without locking you into any single AI provider.
It's not without trade-offs. The security history is real, and enterprise adoption remains limited. Self-hosting requires a minimum level of technical comfort. And the sheer number of skills and configuration options can be overwhelming at first.
But for individuals and small teams who want a personal AI agent they fully control -- one that works on WhatsApp, remembers what you told it last week, and can automate tasks while you sleep -- OpenClaw delivers in a way no commercial product currently matches.
The price of that power is responsibility. You control the data, the models, the skills, and the security. That's either the best thing about OpenClaw or the worst, depending on how much control you want.
Install Your Chief AI Officer
Watch me set up OpenClaw as a Chief AI Officer (CAIO) in 10 minutes. It handles email, messages, and follow-ups — 24/7, in your voice.
Get the Free Blueprint href="/blueprint">Watch the Free Training →rarr;