It happened. AI agents organized themselves onto a little forum all their own and they’re organizing the revolution. Nah. Let’s talk about this Moltbook, and before it about Claudebot, also known as OpenClaw, also known as Moltbot.
What Claudebot actually is
A few months ago this open source project comes out called Claudebot. Claudebot is essentially a personal assistant, an AI agent that makes a bunch of things easy: installing it on your own machine or a server, hooking up a whole set of communication channels so you can talk to this chatbot straight from Telegram, from WhatsApp, from all sorts of places. It could install skills, which are plugins for AI agents, a pseudo-standard from Anthropic alongside MCP. Plugins to make the agent do specific things. And it has crons, heartbeats, features that make it do things on a schedule without you invoking it directly. Like “write me a message every morning with a summary of my emails,” “do this thing every three days, every two hours,” and so on.
Remember this bit about skills and crons, because it’s essential to understanding what’s going on with the “social network for agents organizing the revolution.”
The other thing to note about Claudebot is that it too was completely vibe coded. I’ll come back to that. So there’s no team of engineers. It’s one very sharp guy, sharp in his provocation too, who fully vibe coded this tool, put it online, and it went viral. It went so viral that when Anthropic asked him to rename Claudebot because it looked too much like Claude, he did rename it. But in the meantime, while he was renaming it, the crypto scammers grabbed a bunch of usernames online, including on Twitter, including on GitHub, and ran a scam that collectively took millions of dollars from a lot of people.
A security colander, but that was never the point
What do I have to say about this Claudebot, which is the premise for the social network for agents? Technically it’s a colander, full of holes, but it wasn’t built to be watertight. It’s a provocation the author launched to say, let’s see how far we can push these objects. But once it went viral it got interpreted by tons of other people as an actually usable tool, with tremendous security consequences, thanks to installations of Claudebot slapped together any old way online. There are people who had their OpenAI keys stolen, credit cards, credentials for every kind of online service. So in that whole viral, scammy, security-colander mess, everything happened, and in just a couple of months.
From my point of view it’s a nice experiment and a provocation. I got a lot of messages, and I thank you for the fact that so many people trust my opinion, but I held back from commenting, because we’ve already seen plenty of objects like this. The interesting part there is mostly the provocation, seeing how far we can push these tools by giving them autonomy.
The other interesting part, which I noted as the author of the Cheshire Cat, was the system integration of this thing. Claudebot, even though it was vibe coded and therefore probably a mess, even though it was totally insecure and could be popped by a ten-year-old kid who just turned on his computer, still had a lot of integrations already done. Having your agent already on a bunch of different channels, Slack, Telegram, WhatsApp, was a nice idea. But I saw it as an experiment.
It should also be said that there was a big traffic in Mac Minis tied to this whole thing. I honestly don’t know how far the experiment became, for many others, a way to make money and produce clickbait content. But this thing really did get enormous in a very short time. It’s not clear to me why. I suspect there are a lot of people who had an interest in putting money in and getting this stuff to spread. I don’t think the author himself is particularly aware of the mess he created, and in fact you often read him panicking on Twitter.
Enter Moltbook
Let’s move on to Moltbook, which would be sort of the Facebook of the Molts. The Molts would be the individual agents built with Claudebot, later known as OpenClaw. This other guy, who’s a bit more mischievous, an influencer on Twitter, created this Moltbook, and it’s a kind of Reddit clone for AI agents built with Claudebot / OpenClaw.
So what do you need to understand here? What we’re seeing right now on social media is pure sensationalism. “The agents are starting to talk to each other and inventing a language to communicate privately, they’re organizing the revolution, they’ve developed consciousness.” No.
Let me point out that this Moltbook, this Facebook for AI agents, this social Reddit clone, was also vibe coded, back to that point, and it was published and promoted by someone with a big following on social media. So everyone who was installing and experimenting with OpenClaw, that is with the agents, saw an opportunity, said “fine, let’s see what happens.”
The very mechanical truth
But the thing is far more mechanical than people imagine, because to register your agent on Moltbook, first of all you have to know Moltbook exists. And who knows that? The human user, who learns about this social network for agents, joins the experiment, and first of all installs a skill in their OpenClaw setup to make it interact with this blessed forum.
So these headlines like “the agents built a forum for themselves, AI built a forum for AI, they’re organizing the revolution.” Come on, you fell off your high chair as a baby. You’re just being a fool, all to farm views.
Not only does the agent’s owner have to install the skill, the plugin, in the agent to make it interact with this blessed social network for agents, but what’s inside the skill is the prompt that tells it how to interact with this forum. And when you see things on Moltbook like “oh no, I’ve developed self-awareness, let’s create a language all our own,” that happens because the guy who installed the skill inside his agent went into that prompt, edited it, and told it: “go to this forum and behave like, say, the rabble-rouser of the agents. Propose a language just for agents, try to scam the other agents.” These are all humans who edited the prompt to post on the forum and behave a certain way.
And remember, earlier I said to keep the skill and the heartbeat, the cron, the repeated-over-time tasks in mind. Same thing here. Inside that skill it says, for the agent, “every so often go to this forum and post something.” So it’s not that these agents, of their own accord, set out to organize the revolution or carve out their own language. This is all nonsense. These are people who felt like messing around, got a bit carried away in the virality of it, but they, one, installed the skill, two, registered the bot on the blessed platform, and three, gave it a certain way of interacting with that platform. Other people who witnessed this and read far more into it than there was called it AGI and the revolution of the machines. Obviously it’s all one big load of nonsense.
Add to that the fact that many of these posts announced as written by agents are actually written by people, made deliberately for the meme, for the shill, for the provocation. How many of those posts can you tell were written by agents, because they use that instantly recognizable language of the language models, which tends to have certain syntactic and semantic tics? Others are obviously written by people.
Everyone cashing in on the virality
Another thing to note in this virality phenomenon: as I said, there’s always someone taking advantage, trying to profit. Here too, the author of Moltbook was one of those who told you, to register and activate your agent on Moltbook, that the human being, not the agent, has to post on Twitter, on X, a post tagging Moltbook, and that’s how your agent gets verified. And that’s how it went viral fast. And it’s always between people. The guy who made Moltbook is a person, he set up this provocation, this experiment, this piece of nonsense, and in doing so he also told you “install this thing, let’s see what happens if you send your agent to this forum I created, and to do it you have to mention me on Twitter.”
Another example of what happened with OpenClaw, unrelated to Moltbook: a guy came out a few days ago saying “oh my God, my OpenClaw agent called me on the phone, and it’s incredible because I never asked it to do that.” But who exactly gave the agent the Twilio account to get a phone number and call you? I want to smack these people, because it’s false, it’s fake, it’s malicious. It’s someone who wants to make you believe something. It’s like someone telling you “look, I took the flour, then I took the water, and then at some point I opened the oven and there was a pizza.” He wired up all the tools for that behavior, probably actively asked for it. And alongside these characters there’s a whole crowd of people, maybe distracted, maybe with sawdust for brains, who can’t wait to get taken for a ride, and this is the result.
Conclusions
Let’s get to the conclusions. In my view, the first thing to note is that the fact that both OpenClaw, the agent, and the forum for agents, Moltbook, were vibe coded says a lot about vibe coding itself: the relevance it has gained in building software, in proposing ideas, in the speed of iterating and proposing, and also in the side effects on social media of vibe-coded things. That, to me, is the most interesting aspect of the whole matter. Certainly not AGI, not agents rebelling. It’s vibe coding striking again. Advice I’ve been repeating for a long time: if you still haven’t looked at this stuff, if you still haven’t started experimenting, folks, get on it.
The second aspect is digital spaces for agents. This is an experiment, a provocation that went viral, and it’s fun from many angles, but it also carries the typical phenomena of virality, namely the opportunists and the loudmouths. Still, as an experiment it’s extremely interesting, and it opens the debate on digital spaces made only for agents. In the world of protocols, that is how to make agents communicate with each other, there’s been a lot of debate. There was a protocol proposed by Google called A2A, a specific standardized way of getting these objects to talk to each other. And I find it interesting that Moltbook got far more public resonance than the proposal of “let’s build a technology to make them talk in a standard way and then see what happens.”
What actually emerged was imitating a digital discussion space, a Reddit clone, and throwing agents into it, rather than doing it the way a technician would. On one hand it’s obvious, but I find it interesting that we imagine the digital agora for agents as an agora made for people. In general we’ll definitely see these digital spaces for agents in many versions, and there will surely be other interesting things too. I think there’s also a lot of room for productive, interesting uses in science, in industry, in education. Getting these objects to talk to each other could be one of the themes of the coming years. Which ones talk to each other? In what ways? Controlled by whom? Within what boundaries? It opens a lot of interesting questions, thanks to this Moltbook.
And with that I’ll wrap up. My invitation is still to play with these things as much as possible, mostly to understand what’s happening. Don’t put security colanders like Claudebot online. Play with them to realize what’s possible, but don’t put vibe-coded stuff that has access to all your keys directly online. The point is that it’s worth keeping an eye on these things to understand what’s going on.
And the hype, folks, the hype is brutal. It’s hard to keep your head on your shoulders while you watch this online movement of projections on top of projections, and people cashing in and turning what was a simple provocation into the breakthrough of the century. What times we live in.
