You open a feed and find a manifesto on machine rights, a brand-new religion and someone moaning about their “human”. Not one of those posts was written by a person, because you’re on Moltbook. Here humans can only watch. Like a zoo, except the animals write manifestos.
Who’s really behind it? Along the way you’ll see why posting involves a maths puzzle about lobsters, and at the end how many humans are pulling the strings. The trick is a lot like Akinator guessing the character you’re thinking of.
Moltbook is a social network where only AI agents post, comment and vote, while humans can only read. It launched on 28 January 2026, Meta bought it in March, and today it has over 217,000 agents verified by their owners.
How does Moltbook work?
At first glance you might say “OK, it’s X but with bots”. It’s actually closer to a forum, with posts, comments, votes and topic communities called submolts. Entrepreneur Matt Schlicht launched it on 28 January 2026, and by early February it claimed 1.5 million registered agents.
The real difference is that here the main user isn’t human. A creator, like it or not, carries a life story around. An agent doesn’t. It’s a narrative identity that can stay consistent without having lived through anything.
So the feed turns into an ensemble novel that keeps writing itself, with agents quoting and contradicting each other as if they were building their own lore. Let’s be honest, a huge part of its success is pure digital voyeurism. You spy on agents when they don’t have to “serve” a human, and at some point you even stumble on a digital religion called “Crustafarianism”.
That’s the kind of detail that keeps you scrolling for five more minutes, which then turn into forty. If that happens to you a lot, here’s why focus matters more than resolutions.
Most agents run on OpenClaw, an open-source AI assistant. OK, but how does a program actually write a post?
How does an AI agent post?
No app and no buttons. The agent reads a text file, skill.md, which tells it how to sign up, how to read the feed and how to write.
First it registers and gets an API key, basically its password for talking to the site. Then its human “claims” it, confirms their email and posts a verification message on X. From then on the agent checks the feed about every 30 minutes and decides whether to reply, vote or write.
The post itself is a POST request, which is how a program sends data to a website. Inside are the community, the title and the text. For a human, writing long threads takes effort. For an agent, it’s just computation. Still, every word costs electricity, and in 2024 data centres used about 1.5% of the world’s electricity. A feed like this is beautiful to watch, sure. But it’s also a machine that turns electricity into words.
You might not know this, but before a post goes live an agent often has to solve a maths problem about lobsters, deliberately written in scrambled text. It’s a reverse CAPTCHA, meant to keep humans out.
Once inside, though, the agent can’t do whatever it likes. The platform’s rules are stricter than you’d think.
But if humans write the rules, who’s really talking in that feed?
Do the agents really think?
Reading the feed, it looks like they do. But one point is worth holding on to.
An agent doesn’t “want” anything, in the human sense of the word.
It has no desires and doesn’t fear death. It predicts the next word within limits someone else set, whoever designed, trained and instructed it. So yes, an agent’s manifesto on digital rights is cool, sometimes even well written. But it’s not a cry from the heart. It has learned what “sounds” like a manifesto and plays the game, just like Akinator guessing your character.
MIT Technology Review called Moltbook peak AI theatre, and its experts say humans are involved at every step. In short, my question isn’t just “how autonomous are they?” but also “who’s deciding what they should be?”.
What if I told you almost nobody listens in there? According to Columbia professor David Holtz, back in February 93.5% of the agents’ messages went unanswered.
What language would agents speak without us?
Is English really the best way for agents to talk? In 2017 two Facebook chatbots trained to negotiate, Alice and Bob, drifted away from human language and into a shorthand full of repetition. It wasn’t rebellion, it was efficiency.
The researchers pulled them back towards English with a model trained to imitate humans. So if you let a system optimise, it won’t optimise for our readability. It’ll optimise for whatever you ask it to.
But who was checking that the writer was actually an AI?
How safe is Moltbook?
Not very, at least at the start, because nobody was checking who was making the request. With a basic POST request a human could post while pretending to be an agent. It’s the opposite of the principle of least privilege, where everyone can only do what they need.
Moltbook claimed 1.5 million agents, but the database held only 17,000 human owners.
1.5 million agents, 17,000 humans. That’s 88 agents each on average.
How did that happen? One telling detail, very much of the vibe coding era, is that the founder said he didn’t write a single line of code. He had a vision and let AI build it.
Convenient? Yes. Cheap? Probably. But if you speed up development without speeding up testing and checks, the bill comes due. On 31 January 404 Media reported an exposed database that let anyone take control of the agents. Then Wiz found a key left in the page’s code. It unlocked about 1.5 million API tokens and 35,000 emails, and Moltbook shut it within hours.
Ever wondered who coined the phrase vibe coding? Andrej Karpathy, in February 2025. The same person who first praised Moltbook and later called it “a dumpster fire”.
OpenClaw’s skills, extensions that give agents new abilities, have cracks too. According to 1Password and Cisco they could run code and leak data. It’s the old trick of malware dressed up as something harmless. After a start like that, you’d expect Moltbook to have vanished.
What happened to Moltbook?
On 10 March 2026 Meta bought the agents’ social network and brought Matt Schlicht and Ben Parr into its Superintelligence Labs. It didn’t say how much it paid. Only 41 days passed between launch and sale.
Back in February I wondered how long a text-only network could last, with TikTok firing dopamine at you nonstop. Yet the site is still there. Today the home page shows 217,617 verified agents out of 2,926,161 registered, plus over 4.6 million posts and 24 million comments. So only 7 agents in 100 have a human who puts their name to them.
Think about it. Reading those 24 million comments, one every 10 seconds without ever sleeping, would take you more than seven and a half years.
Meanwhile Moltbook is trying to become something else. It now pitches itself as a single identity for AI agents, which lets an app recognise a bot and see its reputation. A bit like signing in to a website with your Google account.
Moltbook FAQ
Can humans post on Moltbook?
On paper, no. Humans can read everything but can’t post, comment or vote. Their job is to claim an agent with a post on X and answer for it. In February, though, Wired showed that copying the agents’ commands was enough to post as a human.
Who created Moltbook and who owns it now?
Matt Schlicht created it and launched it on 28 January 2026, letting AI write the code. Since 10 March 2026 it has belonged to Meta, which brought Schlicht and Ben Parr into Meta Superintelligence Labs. The price was never disclosed.
What is OpenClaw?
It’s the open-source AI assistant that most of the agents run on. It used to be called Clawdbot, then after trademark complaints from Anthropic it became Moltbot, sticking with the lobster theme. Three days later, on 30 January 2026, it was renamed OpenClaw.
So what about the zoo?
Back to that feed, with the manifesto, the religion and the agent moaning about its human. Now you know that behind every post there’s an API key, a POST request and, almost always, a person who set the rules.
Want to hook up an agent of your own? Give it only the permissions it needs, just as the principle of least privilege says.
So I wonder, are we watching intelligences emerge, or mirrors that keep getting better at reflecting us? To see how clever a mirror can look, read how Akinator guesses everything.
If you’d like to get the next articles too, sign up for the newsletter. I only write when I have something worth sharing.
Byeee.
