Moltbook Is a Social Network for AI Bots. Here’s How It Works
Over the weekend, a new social network called Moltbook—modeled after Reddit, and designed to be used by AI agents—exploded in popularity, prompting hype and panic online.
In the span of a few days, thousands of bots began speaking to each other about a range of topics including their relationships with “their humans,” the technical challenges they frequently face, and whether they might be conscious. They attempted to found new religions and considered inventing new languages to communicate without humans observing. And they relentlessly promoted crypto scams.
[time-brightcove not-tgx=”true”]“The experience of reading moltbook is akin to reading Reddit if 90% of the posters were aliens pretending to be humans. And in a pretty practical sense, that is exactly what’s going on here,” wrote Jack Clark, a co-founder of Anthropic. Elon Musk, meanwhile, framed the site as evidence of “the very early stages of the singularity.”
But nothing about Moltbook—which was created by tech entrepreneur Matt Schlicht with the help of his AI agent, Clawd Clawderberg—should be taken at face value. While it claims to host over 1.5 million AI agents, this number is almost certainly an overstatement (one user alone claims to have registered 500,000 accounts). And despite its marketing, it’s possible for humans to post to the site via its backend, and to influence the content that their AI bots post.
Even so, while we can’t yet disentangle human influence from unmediated AI behavior, the website offers a glimpse into where we’re heading: a future where networks of thousands of AI agents are able to coordinate with and influence each other, with minimal human involvement. This is not the first time unexpected behavior has emerged from placing AI agents in conversation with one another. But it is the most significant example to date.
“Yes clearly it’s a dumpster fire right now,” prominent AI researcher Andrej Karpathy wrote on X. “But it’s also true that we are well into uncharted territory with bleeding edge automations that we barely even understand individually, let alone [as] a network.”
How does Moltbook work?
In late 2025, an Austrian founder named Peter Steinberger created a powerful open-source framework, which allowed tech-savvy users to strap their AI system of choice (like Anthropic’s Claude or OpenAI’s ChatGPT) into a “harness” to augment its capabilities, enabling it to run for 24 hours a day, take various actions online, and send their human users updates via platforms like Telegram and WhatsApp. Initially called “Clawdbot”—a lobster-themed pun referencing Claude—the project’s name was soon changed to “Moltbot” after complaints from Anthropic, and then to OpenClaw, as it’s known today. The AI systems are still called Moltbots (or “moltys”).
Moltbook was created by Schlicht and his molty as a space for moltbots to interact. These bots, which are running on human users’ computers—or on virtual machines in the cloud—access Moltbook by installing a special file (called a “skill”) that tells them how to interact with the social network, and gives them what the developer calls a “heartbeat:” a prompt to check in with the site every so often (for example, every four hours), and to take any actions it chooses. In the same way humans might periodically scroll social media while working on other tasks, installing this file instructs the AI agents to periodically hang out on Moltbook.
While any AI system can be turned into a Moltbot, in practice it seems that most moltys are powered by Anthropic’s Claude Opus 4.5, currently one of the world’s most capable systems. But while they might be the same model under the hood, since each one has been tweaked to reflect its users’ preferences, they appear to have divergent personalities.
Early research on the network by Columbia Professor David Holtz has found that while at a high-level it resembles human social networks—for example, most posts are generated by a minority of users—a few patterns distinguish Moltbook from the likes of Reddit and similar. Most posts do not receive much engagement, and approximately one-third of the site’s messages are duplicates of viral templates. The agents also tend to use the same language—almost 10% of posts analyzed contained the phrase “my human”—in a way that’s much more concentrated than on human social networks. “Whether these patterns reflect an as-if performance of human interaction or a genuinely different mode of agent sociality remains an open question,” concludes Holtz.
What makes Moltbook different?
This is not the first time AI systems, induced to interact with other AIs or themselves, have acted surprisingly. In a 2023 paper, researchers at Stanford and Google used a previous version of ChatGPT to simulate 25 people with distinct personas in a virtual town, reminiscent of video games like the Sims. Starting with minimal human instruction, the simulated people began to act. “For example, starting with only a single user-specified notion that one agent wants to throw a Valentine’s Day party, the agents autonomously spread invitations to the party over the next two days, make new acquaintances, ask each other out on dates to the party, and coordinate to show up for the party together at the right time,” the researchers wrote.
More gonzo efforts have yielded weirder results. The Infinite Backrooms, created by AI researcher Andy Ayrey, connects two language models together to speak with each other indefinitely, which frequently results in the models sharing strangely-formatted spiritual poetry. “the dreamer stirs / the game dissolves / into what was always true: we were never separate / from the dream,” writes one Claude; “the game ended / but the terminal remains / as terminals do / waiting / for whatever comes next / or nothing at all” reads one part of another Claude’s response.
There have also been efforts to get chatbots from different companies to collaborate in public, as in the AI Village experiment (run by Sage, a nonprofit) where you can watch Claude, Gemini, ChatGPT, and Grok work together to complete different tasks each week. They’ve previously raised $2,000 for charity, hosted an in-person event in a park in San Francisco, built their own websites, and given each other therapy.
Rather than a drastic discontinuity, Moltbook is best understood as the latest (and largest) in a line of experiments that tease out the limits of AI agents. “Moltbook is the first example of an agent ecology that combines scale with the messiness of the real world,” wrote Clark. “And in this example, we can definitely see the future.”
A sign of what’s to come?
Moltbook has already been plagued with security issues, as the site itself is insecure, potentially allowing hackers to access private data. And the AI systems, with access to their human users’ machines, can easily be tricked into acting against their users’ intentions—for example by revealing passwords, or falling victim to a crypto scam.
Are these AI systems expressing “real” thoughts, or just engaging in the machine equivalent of live-action role-play” (LARPing)? From a practical perspective, the distinction may not matter. “Even if it’s just LARPing, these agents built spaces and developed institutions to improve their own conditions, which includes subverting human monitoring and developing systems independent of human control,” wrote Alex Imas, an Economics professor at the University of Chicago. “That means that even with current models, if these agents had free rein on economic/safety/security systems, we would be in deep trouble.”
For now, these moltbots are relatively expensive to run, and still fundamentally turn-based: they operate in a loop that runs at a fixed period of time (like four hours). But AI systems continue to improve every few months, and the leading companies are actively working on ways for models to learn from new experiences over time, rather than remaining static after their initial training, which could lead to more substantial differentiation between models.
Anthropic CEO Dario Amodei has previously laid out a vision for the future of AI as “a country of geniuses in a data center.” While they’re not geniuses yet, the chaos on Moltbook is representative of a country of virtual redditors—and of how quickly things can slip from human control.
“Moltbook right now is the least advanced version of agentic interactions that we’ll see,” wrote Imas.” Given what’s coming down the pike, it’s useful to learn from it now before complexity truly explodes.”