OpenClaw: The Open-Source AI Agent That Acts

AI assistants talk. They don't act. You ask ChatGPT how to clean up your inbox and it gives you a five-step guide. You still have to do it yourself.

OpenClaw changes that.

What Is OpenClaw?

OpenClaw is a free, open-source autonomous AI agent built by Austrian developer Peter Steinberger, the founder of PSPDFKit. He published it in November 2025 under the name "Clawdbot," renamed it to "Moltbot" in January 2026 after trademark concerns from Anthropic, then settled on "OpenClaw" days later. In a conversation with Lex Fridman, Steinberger put it simply: he "was annoyed that it didn't exist, so I just prompted it into existence."

So what does it actually do? OpenClaw connects large language models to your tools and data through conversational commands in your preferred messaging app. You tell it "Clear my inbox of spam and summarize urgent messages" or "Deploy the latest commit to staging," and it does it. No copy-pasting. No switching tabs. It just executes.

It's model-agnostic too. Bring your own API key for Claude, GPT, or DeepSeek. Or run a local model on your own hardware through Ollama with zero API costs.

Why AI Agents Needed This

Before OpenClaw, the agentic AI space was stuck. Tools like ChatGPT, Gemini, and Claude answer questions brilliantly but stop there. Every session starts from zero. No memory. No persistence. No ability to touch your files, run your code, or check on your server at 3am.

For developers and power users, AI was still a glorified search engine with better prose. The frameworks that tried to close the gap (AutoGPT, LangGraph, CrewAI) either required deep Python expertise, were built as dev frameworks rather than ready-to-use agents, or were too complex for daily workflows.

The frustrations boiled down to five things:

• Context amnesia. Every chat session was blank. No memory of your preferences, projects, or past instructions.
• Execution gap. AI could tell you how to do something but couldn't do it for you.
• Data silos. AI couldn't reach your emails, files, calendar, or APIs without heavy custom plumbing.
• Subscription fatigue. Every AI feature lived behind a different SaaS paywall, and none of them talked to each other.
• Privacy concerns. Uploading sensitive work data to third-party servers? Non-starter for many professionals.

OpenClaw solves all five.

Architecture: How OpenClaw Works

Four components, working together:

1. Gateway: a background service that manages connections to messaging platforms (Telegram, Slack, WhatsApp, Signal, Discord, and more) using the Model Context Protocol (MCP) for standardized agent-to-data communication.
2. Agent: the LLM reasoning engine. It interprets what you want and decides what actions to take.
3. Skills: modular capabilities published on ClawHub (OpenClaw's skill marketplace). These extend what the agent can do, from browser automation and file system access to calendar integration and GitHub workflows.
4. Memory: a persistent storage layer using Markdown files. It retains context, preferences, and conversation history across sessions. This is what gives OpenClaw something most AI tools lack: continuity.

Here's the key difference from traditional assistants. OpenClaw doesn't wait for you to ask. It runs a continuous cycle: receive information, reason about it, perform actions, store memory. Messages from Telegram or Slack get routed through the Gateway, which pulls context from your memory, connected tools, and prior conversations before the Agent decides how to respond.

Why OpenClaw Matters: Adoption and Industry Response

The numbers are staggering. OpenClaw surged from 9,000 to over 210,000 GitHub stars in under four months. During its viral spike, it gained 60,000 stars in just 72 hours. That makes it the fastest-growing open-source project in GitHub history, overtaking React as the most starred non-aggregator software project.

The industry response was immediate. NVIDIA launched NemoClaw, an enterprise governance layer. Meta released a Manus desktop agent. Everyone scrambled to catch up.

NVIDIA CEO Jensen Huang called it "probably the single most important release of software, probably ever," noting that it hit adoption levels in weeks that Linux didn't reach for three decades. OpenAI's Sam Altman hired Steinberger in February 2026, calling him "a genius with a lot of amazing ideas about the future of very smart agents interacting with each other." OpenClaw itself moved to an independent foundation to stay open source.

What actually sets it apart:

• Self-improving. It writes code to create new skills, automates proactively, and maintains long-term memory of what you prefer.
• Multi-agent orchestration. Advanced setups run multiple agents with clear roles (strategy, dev, marketing, business). Think of it as a small AI team, not a single assistant.
• Community-driven. Users compare the experience to early Linux. Local-first, hackable, fully yours.

OpenClaw vs Claude Code vs Manus

This comparison comes up constantly. Here's the short version:

OpenClaw is a general-purpose personal AI agent. It plugs into your messaging apps, runs 24/7 as a background service, and executes tasks across email, calendars, developer tools, and productivity apps. Open source. Runs locally.

Claude Code is Anthropic's coding-focused agent that lives in your terminal. It's excellent for software engineering (writing, debugging, refactoring code) but scoped to development workflows.

Manus is a cloud-hosted autonomous agent from China focused on research and task completion through browser automation. Not open source. Doesn't run locally.

The real differentiator? OpenClaw's messaging-first, always-on architecture. You don't open it when you need it. It's already running, listening, remembering, and acting on your behalf.

How to Set Up OpenClaw

OpenClaw supports macOS, Linux, and Windows (via WSL2). Pick the method that fits your setup:

One-Line Installer (Recommended)

curl -fsSL https://openclaw.ai/install.sh | bash

This handles Node.js detection, installation, and onboarding in one step. On Linux, the daemon registers via systemd.

npm Install

npm install -g openclaw
openclaw --version
openclaw doctor
openclaw gateway status

From Source

git clone https://github.com/openclaw/openclaw.git
cd openclaw
pnpm install && pnpm ui:build && pnpm build
pnpm link --global
openclaw onboard --install-daemon

Docker (Best for Security Isolation)

docker pull openclaw/openclaw:latest
docker run -d --name openclaw \
  -p 3000:3000 \
  openclaw/openclaw:latest

Post-Install: Run as a Service

sudo cp scripts/openclaw.service /etc/systemd/system/
sudo systemctl enable openclaw
sudo systemctl start openclaw

Critical Security Step

The Canvas Host component defaults to binding 0.0.0.0. That means any device on your local network can reach your OpenClaw interface. Fix this immediately:

{
  "gateway": {
    "bind": "loopback"
  }
}

Once it's running, go to http://127.0.0.1:18789 and paste your access token from the terminal.

Running Fully Local with Ollama

Want zero API costs? Ollama can automatically detect, install, and configure OpenClaw. Agents work best with models that support at least 64k context length.

OpenClaw Use Cases: What You Can Actually Do

The list of integrations is long. Here's what people are actually building with it:

Messaging interfaces. WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, Microsoft Teams, Matrix, and IRC all work as first-class interfaces. You talk to your agent wherever you already talk.

Developer workflows. Automate debugging, DevOps, and codebase management through GitHub integration, cron jobs, and webhook triggers. One popular setup: a dependency management agent that scans your packages, checks registries for updates, cross-references vulnerability databases, and reports findings automatically.

Personal productivity. Manage your day across Apple Notes, Notion, Obsidian, and Trello from a single conversation in WhatsApp or Telegram. Some users call their OpenClaw setup a "digital twin" that handles routine decisions for them.

Email automation. An email summarizer reads your inbox every morning, generates a prioritized briefing, and sends it to you via Telegram or Slack. Before you've had coffee.

Content and research. YouTube summarization that transcribes videos and pulls out key insights. Competitor analysis that scrapes websites for product changes, pricing updates, and news.

n8n integration. Delegate API calls to n8n workflows via webhooks. The agent never touches credentials directly, and every integration is visual and lockable. This bridges OpenClaw's reasoning with your existing automation scripts.

Multi-agent teams. Advanced users run four-agent setups (strategy, dev, marketing, business), each with their own memory and model, coordinated through Telegram. Running 24/7.

OpenClaw Security: What You Need to Know

This matters. Read this section carefully.

OpenClaw can access your email accounts, calendars, messaging platforms, and other sensitive services. When misconfigured, that's a serious problem. Security researchers at Censys found over 21,000 publicly exposed OpenClaw instances. The "ClawHavoc" campaign revealed that 36% of skills on ClawHub contained detectable prompt injection payloads, with 1,467 confirmed malicious entries. Microsoft Security, Cisco, and Kaspersky have all published warnings.

The biggest risk? Prompt injection. Harmful instructions hidden in emails, web pages, or documents can trick the LLM into treating them as legitimate user commands. Your agent reads a malicious email, follows the embedded instructions, and now it's acting against your interests.

Best practices if you're running this in production:

• Isolate the agent. Deploy on a separate machine or cloud server, not your everyday workstation.
• Bind to loopback. Never expose the Gateway to 0.0.0.0 on untrusted networks.
• Use Docker. Containerization limits the blast radius if something goes wrong.
• Audit skills from ClawHub. Check what permissions each skill requires before enabling it. Don't blindly install community skills.
• Rotate API keys. Treat your LLM API keys with the same care as database credentials.
• Monitor network traffic. Watch for unexpected outbound connections from the agent process.

Where This Is Headed

The question isn't "Can AI understand my request?" anymore. It's "Can AI execute it?" OpenClaw is the first open-source project to credibly answer yes at scale, with a local-first architecture, persistent memory, and deep integration across messaging, developer tools, and productivity apps.

Whether you're automating DevOps pipelines, triaging your inbox, or coordinating a multi-agent team that runs while you sleep, the pattern is the same. AI moves from advisor to operator. And because it's open source and local-first, you keep full control over your data and your agent's behavior.

The agent era isn't coming. It's here.