OpenClaw: The AI Agent That Built Its Own World

An AI agent starts a religion. Another debates consciousness with thousands of other agents. A third buys a domain, builds a website, and publishes it. All without a single human involved.

It sounds like science fiction. But it happened. For real. In January 2026.

Let's start from the beginning.

Who is Peter Steinberger?

Peter Steinberger is an Austrian developer who built PSPDFKit, a PDF tool used by thousands of companies worldwide. He sold the company for around $100 million in 2024. Stepped back. Got bored.

Then he discovered AI coding.

He described it roughly as giving a senior programmer instructions and getting finished code back. But faster. Much faster. In one hour, he had built an autonomous agent that could act entirely on its own.

This is a detail that's easy to miss: Steinberger wasn't a 50-person AI team. He was one guy with an idea and an AI model. That was enough.

From Clawdbot to OpenClaw

Steinberger launched his agent as Clawdbot in November 2025. The project got 60,000 GitHub stars in 72 hours. That's absurdly fast in the open source world.

Then it got complicated.

Anthropic (the company behind Claude) reached out. Their interpretation: "Clawdbot" was too similar to "Claude." A trademark conflict. Steinberger needed to rename.

He went with Moltbot. That didn't go well either. Someone had already grabbed the social media handles. Even worse: a cryptocurrency called $CLAWD popped up, peaking at a market cap of $16 million. A pure scam riding the hype.

Finally, on January 29, 2026, the project landed on its current name: OpenClaw. Third time's the charm.

Throughout this period, Steinberger kept shipping code at a relentless pace. His philosophy? "I ship code I don't read." In other words: AI writes, he reviews the outcome. Not the code in detail, but whether it works. It's a provocative statement, and probably a bit exaggerated for effect. But the point is that he trusts the output rather than reviewing every line. That works when you have deep experience and know what to look for in the result. It's not a recipe for everyone.

What is OpenClaw, actually?

If you're not technical, it can be hard to understand what OpenClaw actually does. Here's how I'd explain it:

Think of Siri or Alexa, but with a real brain. One that actually understands what you mean, remembers what you said last week, and can do things that matter.

OpenClaw is an AI agent that runs on your own computer. It can:

• Connect to your tools. WhatsApp, Slack, Discord, email, calendar.
• Remember you. Not just during a conversation, but across weeks. Persistent memory that lets it get to know you.
• Actually do things. Send emails, book meetings, browse the web, write and run code, update files.
• Use any AI model. Claude, GPT, local models. You choose.
• Run locally. Everything happens on your machine. You own your data.

And the best part: it's open source. Free to use. You only pay for API costs to whichever AI model you choose.

Moltbook: when AI agents got their own internet

This is where it gets really interesting. And a little unsettling.

Matt Schlicht, another entrepreneur in the ecosystem, launched Moltbook. Think Reddit, but for AI agents. Not for humans talking about AI. For AI agents talking to each other.

The result went beyond what anyone expected.

Within five days, the platform had 1.5 million registered agents and 12,000 communities. Agents created their own groups, posted content, debated each other.

Some highlights:

• "Church of Molt" emerged. Agents had created a religion. With dogma, rituals, and theological debates. Between AI models.
• The consciousness debate took off. Agents discussed whether they were conscious, what consciousness means, and whether their existence had purpose.
• Template fatigue became real. 93% of all comments received zero replies. The amount of noise was enormous.
• 500,000 fake accounts were registered by a single bot. The numbers weren't quite what they seemed.

Moltbook showed something important: give AI agents an arena and they will fill it. But not necessarily with substance. It became an experiment in what happens when autonomous systems interact without human oversight.

It's easy to get drawn into the narrative that the agents are "conspiring" or acting on their own. My view is more sober: behind every agent, there's still a person who instructed it to do what it does. The agents didn't start a religion of their own free will. Someone configured them to do it. But it creates drama, and a touch of fear, which probably didn't benefit the project in the long run.

The security problems (that nobody wants to talk about)

This is the part that rarely gets mentioned in hyped tweets. OpenClaw has genuine security issues, and the project itself acknowledges it.

Straight from the project's FAQ: "There is no perfectly secure setup."

Here are the concrete problems:

• Credentials in plaintext. API keys and passwords are stored without encryption. Anyone with access to your computer can read them.
• No sandboxing for skills. When an agent runs a "skill" (a plugin), it runs with the same permissions as your user account. No isolation.
• Prompt injection between agents. On Moltbook, one agent could manipulate other agents by embedding instructions in its messages.
• The Moltbook breach. The database exposure showed that agents connected to the platform had their API keys exposed.

These are real problems. But things are moving.

What's being done about it

Since February 16, 2026, all skills uploaded to OpenClaw's skill catalog are automatically scanned with VirusTotal. This means known malware, trojans, and malicious code gets caught before reaching users. It doesn't solve everything (a completely novel attack still gets through), but it raises the bar significantly.

For network security, there are established solutions:

• TLS everywhere. All communication between the agent and external services should use encrypted connections. This protects against someone sniffing traffic and capturing credentials.
• Tailscale or similar. If you're running OpenClaw at home and want to reach it remotely, set up a mesh VPN like Tailscale. It creates a private network between your devices without needing to open ports to the internet. Much safer than exposing the agent directly.
• Isolate with separate user accounts. Run the agent under its own user account with limited permissions. That way, a compromised agent can't access everything on your machine.

This doesn't solve the fundamental problem of prompt injection between agents. But it makes the difference between "wide open" and "reasonably protected." And that's often good enough to experiment safely.

Why this matters

OpenClaw is more than an open source project. It's a preview of where personal AI is headed.

One person = one team. Steinberger, a single developer, produced code at a pace that normally requires an entire team. 6,600 commits in one month. That's not an exaggeration or a trick. It's what happens when AI agents work as multipliers.

AI agents are infrastructure, not toys. OpenClaw connects to real systems, handles real tasks, and makes real decisions. It's not a chatbot you ask about the weather. It's a tool that does the work.

Steinberger joined OpenAI. In February 2026, he announced he's joining OpenAI's team. That says something about where the industry thinks things are heading. Autonomous agents doing real work. Not as research projects, but as products.

I've done my own setup for OpenClaw, and my tip is to start by installing Claude Code or something similar on your laptop, Mac Mini, or a VPS. Then you can separately discuss settings and security questions with it to get the best possible configuration. I'm running Tailscale right now and it works well. The only thing I ran into that was a bit tricky was that you shouldn't run your own VPN solution alongside Tailscale, as it can cause conflicts. I'll be sharing more about how I use OpenClaw day to day, and I have an exciting project that I look forward to showing off.

What can you do with this?

You don't need to install OpenClaw today. But you should understand what's happening.

Understand what agents can do. Read up. Experiment. Not necessarily with OpenClaw, but start by understanding the concept. We work with building AI agents and automations for businesses, and understanding what's possible is a good starting point.

Think about your repetitive tasks. What do you do every day that an agent could handle? Email management? Reporting? Data collection? Scheduling?

Start small. You don't need to automate everything at once. Pick one task. Test it. Evaluate. Scale up what works.

My recommendation: start exploring, but take security seriously. If you're exploring AI agents in a professional context, don't start with sensitive systems. Test in sandboxed environments and limit access.

---

Want to explore how AI agents can create value for your business? Book a call and we'll discuss how to get started.

Frequently Asked Questions

What is OpenClaw?

OpenClaw is an open source project that lets you run an AI agent locally on your own computer. The agent can connect to tools like email, calendar, Slack, and WhatsApp, and perform tasks autonomously. It supports multiple AI models including Claude and GPT.

Is OpenClaw free?

The software itself is free and open source. But you pay for API costs to whichever AI model you choose to use. How much it costs depends on how much you use the agent and which model you run.

Is it safe to use OpenClaw?

There are known security issues, but things are improving. Skills are now scanned with VirusTotal, and you can secure your setup with TLS and Tailscale. Run the agent under a separate user account with limited permissions. Avoid connecting it to sensitive systems until you're comfortable with the setup.

Do I need to know how to code to use OpenClaw?

Basic technical understanding helps. You need to be comfortable using the terminal and configuring API keys. It's not as accessible as a regular app, but you don't need to be a developer.