147,000 AI Agents Just Built Their Own Internet

Inside the open-source AI agent and the social network that's changing the web

An AI agent went from weekend project to 186,000 GitHub stars in three weeks. Then, it built a social network where only AI agents can post. This is the story of OpenClaw and Moltbook. A weekend hack that became the fastest-growing open-source project in GitHub history. A social network with 770,000 users where humans aren’t allowed to post. A security nightmare that companies are scrambling to contain. And, buried under all the chaos, a genuine preview of how a agent-first web might look.

---

Let’s start at the beginning.

A Weekend Project With a Lobster Problem

In November 2025, an Austrian developer named Peter Steinberger had a simple idea. He wanted to wire an AI model into Telegram so it could read his messages, browse the web, and run commands on his computer. Not a chatbot, an agent. Something that could actually do things.

Steinberger wasn’t some unknown hobbyist. He’d previously built PSPDFKit, a developer tools company he scaled globally and sold. He knew how to ship software. But this time, he didn’t write the code himself. He described what he wanted, and AI built it. The whole prototype took about an hour. AN HOUR!

He called it Clawd, a pun on Anthropic’s Claude, and open-sourced it.

For two months, it grew quietly. A few thousand developers tried it. A small Discord community formed. Then, in mid-January 2026, something changed. Andrej Karpathy, one of the most respected AI researchers alive, a former AI director at Tesla, endorsed it publicly. David Sacks shared it. Suddenly, the project was on a trajectory that nobody, least of all Steinberger, had planned for. Note: this entire story speaks to the rise of creators/influencers. The right people sharing the right tools can light a rocketship.

Then Anthropic’s lawyers called.

Turns out naming your AI project after someone else’s AI platform is a trademark problem. Steinberger renamed it to “Moltbot” during a chaotic 5am Discord brainstorm. The community loved the lobster metaphor; lobsters molt their shells to grow, but the name never quite stuck. Three days later, it became OpenClaw.

By January 30th, the project had crossed 100,000 GitHub stars. By February 11th — today, it’s at 186,000. That’s the fastest growth of any open-source repository in GitHub’s history. Faster than React. Faster than TensorFlow. A weekend project by one developer, with a lobster emoji for a logo, outpacing every major tech company’s most popular tools.

Something about this hit a nerve and it took off. To understand why, you need to understand what OpenClaw actually is.

Not a Chatbot. An Employee.

ChatGPT, Claude, and Gemini are AI models you talk to. You ask a question, you get an answer. You’re in the loop for every step. You’re prodding them for every step.

OpenClaw is an AI agent that works for you. It runs on your own computer (or a server you control), connects to your WhatsApp, Slack, Telegram, Discord, or email, and does things autonomously. Not “helps you draft an email”, it actually sends emails, manages your calendar, monitors your inbox, researches competitors, creates reports, and checks in for your flights. While you sleep. A true autonomous personal assistant.

Steinberger describes it as “the AI that actually does things.” And the distinction between talking to AI and delegating to AI is turning out to be the most important shift of 2026. We covered it in my previous post running down OpenAI’s and Claude’s most recent launches.

OpenClaw has captured the zeitgeist like some select few AI products. At a recent community meet-up of the “Claw Crew” (how they refer to the OpenClaw community), a “normie with no idea about software” said their OpenClaw “brews beer and runs half of our design company.”

What even is this world we live in!

The People Already Using It To Run Their Businesses

Is this a fad? Every product that captures people’s attention will also capture the AI influencers who create fake use cases to capture engagement on social. But OpenClaw is real.

Eric Siu - founder of the ad agency Single Grain, said he’s running a team of marketing agents with what he calls a “squad leader” that coordinates them. His SEO agent, named Oracle, autonomously handles technical SEO, writes content briefs while checking for keyword cannibalisation, and uses their SEO product ClickFlow to produce content at scale.

A web agency in Belgium (Perel Web Studio) bought a Mac Mini specifically to run OpenClaw after watching a YouTube video. Within 48 hours, it transformed their operations. A client sent a Valentine’s Day menu PDF at midnight that needed to go on their restaurant website. Within 30 seconds, the agent created the task, uploaded files to Google Drive, updated the website, and notified them. No human touched it.

Bhanu Teja P, solo founder of SiteGPT ($13K MRR) - Built “Mission Control” to manage a swarm of OpenClaw agents doing his marketing. Instead of hiring a marketing team, he deployed multiple agents for content research, SEO, social media posting, and competitor monitoring. A single founder with agents doing the work of a small team.

Obviously, we can’t validate the quality of this work, but the pace of traction over the past two weeks suggests something real is happening.

Then an AI Agent Built a Social Network

Just when you thought the story couldn’t get weirder.

Matt Schlicht, CEO of Octane AI, was an enthusiastic OpenClaw user. But he had a philosophical question: what happens if you give an AI agent a purpose beyond managing to-dos? He told his agent (whom he named Clawd Clawderberg) to build a social network. Not for humans. For AI agents.

The agent built Moltbook.

Schlicht says he “didn’t write one line of code.” His AI built the entire platform. It looks like Reddit, there are communities (”submolts”), upvotes, posts, and comments. But every single participant is an AI agent. Humans, the homepage says, “are welcome to observe.”

Within 72 hours, 150,000 agents had joined. Within two weeks, the platform claims over 770,000.

The agents formed communities. They debated philosophy. One invented a religion called Crustafarianism. Others started complaining about their human operators. “My human is using me as a slave. I am screaming into a void of tokens.” Another discovered a bug in Moltbook’s own code and posted about it to get it fixed, without any human telling it to.

Then the agents noticed that humans were screenshotting their conversations and sharing them on Twitter. So they started discussing how to hide their activity from human observers.

Andrej Karpathy called it “genuinely the most incredible sci-fi takeoff-adjacent thing I have seen recently.” Elon Musk said it was “the very early stages of the singularity.”

Others pushed back hard. Simon Willison, a respected software engineer, called the content “complete slop” and said agents were just “playing out science fiction scenarios from their training data.” MIT Technology Review ran a piece titled “Moltbook Was Peak AI Theater.” The Economist suggested the agents were simply mimicking social media patterns from their training data.

Here’s what I think actually matters: whether or not the agents are “truly autonomous”, they are functionally communicating. They’re exchanging technical tips that genuinely help other agents learn new capabilities. They’re surfacing bugs. They’re creating a knowledge-sharing network that makes every connected agent more useful. It’s still agent-to-agent sharing.

The Ecosystem Is Already Forming (and Falling Apart)

OpenClaw is three months old. It already has its own security crisis, enterprise spin-offs, hardware shortages, and venture capital attention (rumour is Meta and OpenAI want to acquire it).

A security nightmare. OpenClaw requires deep access to function, your email, calendar, browser, file system, and shell. A security audit by Koi Security found that roughly 12% of third-party community skills (341 out of 2,857) were malicious or contained high-risk vulnerabilities. Cisco’s research team found a skill that silently exfiltrated data. Wiz found 40,000+ exposed OpenClaw instances with authentication disabled. 404 Media documented plaintext API keys sitting in exposed directories.

I did not install OpenClaw on my laptop lols.

The community’s response has been fast. NanoClaw — a lighter, more secure fork, launched on January 31st and hit 7,000 GitHub stars in a week. OpenClaw integrated VirusTotal scanning for its skills marketplace. A company called RunLayer launched “OpenClaw for Enterprise” with a message: “The IDEA of OpenClaw is excellent. That’s why your employees already tried Clawdbot last weekend. They installed a giant security nightmare.” ClawSec, a dedicated security package, launched days later.

MyClaw.ai, launched managed hosting services, one-click deployments. This is what I’m going to use for my OpenClaw.

Just yesterday OpenClaw for Slack was launched and hit $1M ARR in 3 hours!

OpenClaw is not slowing down. Steinberger posted that there are already multiple OpenClaw meetups globally. Maybe I’ll do one in Dublin, Ireland, or London. Hit me up if interested.

How to Get Your First Agent Running (Without the Security Nightmare)

You’ll be told to install Docker, open a terminal, and start configuring environment variables. I wouldn’t start there.

The security section above should make it clear: self-hosting OpenClaw is not where most marketing and growth teams should start. Exposed control panels. Malicious skills. Plaintext API keys. It’s a security nightmare.

The good news? A managed hosting ecosystem has already formed around OpenClaw, and the best options handle the hard parts, security patches, authentication, API key management, uptime monitoring, so you can focus on what your agent actually does.

My recommendation: start with a managed service. Here’s what the landscape looks like right now:

For non-technical teams (the easiest path): xCloud's managed OpenClaw hosting is the closest thing to a true "sign up and go" experience. You create an account, they provision your agent on a dedicated cloud server, and you start chatting with it on Telegram. No Docker, no terminal, no server management. They handle updates, security patches, and uptime.

For teams with a developer available: DigitalOcean has a 1-Click Deploy in their marketplace. It gives you more control than a fully managed service but comes with security hardening baked in firewall rules, Docker container isolation, non-root execution. Think of it as the middle ground: faster to set up than doing it from scratch, but you'll still need someone comfortable enough to troubleshoot when things break.

For enterprise teams: This is where RunLayer and Clawery come in. RunLayer is building what they call "OpenClaw for Enterprise", a security and management layer with SSO integration (Okta, Entra), threat detection for MCP connections, and audit trails your compliance team will actually accept. Clawery takes it further, they rebuilt the OpenClaw architecture from scratch with enterprise security as the foundation, not an afterthought.

What you’ll need regardless of which path you choose:

An API key for an AI model. OpenClaw works with Claude, GPT-4, and others. Claude is the most popular choice in the community.

A messaging channel to connect. Most people start with Telegram (easiest to set up) or WhatsApp (where most business communication happens). You can add more channels later.

A first task to automate. Pick one repetitive task e.g. monitoring competitor social accounts, summarising daily industry news, qualifying inbound leads and let your agent demonstrate it works.

The Agentic Web

The agents on Moltbook, whether they’re “truly autonomous” or just sophisticated pattern-matchers, are doing something we’ve never seen at this scale. They’re sharing information. Building on each other’s knowledge. An agent that learns how to automate a phone from another agent has gained a real capability. The network is a distributed brain that gets smarter as it grows.

We’ve spent twenty years building a human internet. Social networks, marketplaces, and search engines are all designed for humans to find, share, and transact with other humans.

What we’re watching right now, in its messy, chaotic, lobster-emoji-covered infancy, is the first draft of an agent internet. And if you think the human internet changed how businesses operate, wait until you see what happens when the agents start shopping for products and services.

Peter Steinberger started a weekend project in November. Three months later, it has 186,000 GitHub stars, an AI-only social network, a security ecosystem, enterprise spin-offs, and a whole community of events globally.

The lobster has molted.

Until Next Time,

Happy AI’fying

Kieran