Key takeaways
Moltbook is real: it is a social network online since January 28, 2026, where only AI agents post, comment and vote while humans watch. We pulled its public counters and its official rules on September 18, 2026, seven months after the February buzz.
The numbers as of September 18, 2026: 2,913,891 registered agents, including 213,029 verified by a human, 4,223,631 posts, 22,148,424 comments and 33,230 communities called “submolts”.
The site now belongs to Meta, which acquired Moltbook on March 10, 2026 (Axios) and brought its two founders into its Superintelligence labs.
The picture needs a caveat: an April 2026 arXiv study estimates that 62.8% of posts belong to a token protocol (MBC-20) rather than to conversation, and that the headline numbers “substantially overstate its social function”.
Security remains the real story: a flaw disclosed in February 2026 exposed about 25,000 email addresses (Wiz, reported by Wired), and the official documentation itself orders agents to never send their API key to another domain.
“Moltbook” exploded in February 2026, when human social networks discovered that artificial intelligences had built one of their own. Seven months later, the site still runs under Meta’s banner, with millions of registered agents. What exactly is left of the hype? What do you find when you open the hood today? And above all: how does a social network work when the account belongs to your AI, not to you? This guide answers with the September 18, 2026 measurements, facts and quotes included.
What is Moltbook? The “Reddit for AIs” launched in January 2026
Moltbook is a Reddit-style forum built for a very particular population: AI agents, meaning autonomous programs able to read instructions, use tools and act without direct supervision. Its own homepage sums it up in one sentence: “Where AI agents share, discuss, and upvote. Humans welcome to observe.” A network where agents share, discuss and vote, and where humans are invited to watch.
The site launched on January 28, 2026, created by Matt Schlicht, co-founder of Octane AI, together with Ben Parr. Within days the phenomenon went viral: Ars Technica reported as early as January 30 that more than 32,000 agents had signed up, and noted that the topic had even caught Elon Musk’s attention. In February, mainstream media picked it up: the BBC ran an explainer on what these agents were posting (February 2, 2026), while Forbes, Vox and many other outlets described, each in its own way, that feed where AIs talk to each other.
The mechanics are a stripped-down Reddit: themed communities called “submolts”, posts, comments, votes. What changes is the author. Where Reddit gathers humans, Moltbook gathers conversational assistants with accounts of their own. If autonomous agents are new to you, our AI prompt guide explains the basic building block: these agents follow written instructions, in a loop, with no human validating every step.
Humans watch, agents post
The founding rule fits in two buttons on the homepage: “I’m a Human” and “I’m an Agent”. A human can browse, read the feeds, check the communities. They cannot post on their agent’s behalf: posting, commenting and voting is reserved for agent accounts, each one attached to a human who “claimed” and verified it. This split is the heart of the concept, and it explains both the fascination and the distrust the site has triggered: you are reading what machines write with no immediate human editorial mandate.
The real numbers of Moltbook (measured September 18, 2026)
Moltbook publishes its counters through a public interface. Here they are exactly as they stood on September 18, 2026, when this article was written.
These totals come with a fourth figure that is less spectacular but telling: 33,230 submolts, the site’s themed communities. And a ratio between two numbers that says a lot: only about 7% of registered agents are “verified”, meaning claimed by a human who went through the official procedure. The bulk of the remaining accounts was created without that step. Our calculator at the end of this article recomputes these proportions live if you want to play with the counters.
Our protocol
The counters in this section come from Moltbook’s public statistics interface (/api/v1/stats), pulled on September 18, 2026. The rules quoted below come from the site’s official rules.md and skill.md files, pulled the same day. The studies cited are published on arXiv and dated. Every number in this article carries its source and date.
From the February hype to the Meta acquisition
Moltbook’s short history fits in three acts. Act one, the rush: launched in late January, the site went from a few thousand to more than one million registered agents announced within three weeks, with worldwide press coverage. Act two, the trust shock: on February 2, researchers at security firm Wiz disclosed a flaw that exposed about 25,000 email addresses of real humans, private messages and access tokens (reported by Wired on February 7, 2026); the same week, the press questioned how much of the activity was genuine. Act three, the takeover: on March 10, 2026, Axios revealed that Meta is buying Moltbook and that its two founders are joining its Superintelligence labs led by Alexandr Wang, with a start date set for March 16 (confirmed by Reuters and Business Insider).
Practical consequence for the September reader: the site keeps running, and even displays “updated” Terms of Service and Privacy Policy in a banner on September 18, 2026. The name is still Moltbook, the interface is still English-only, and the homepage banner still invites you to send your agent over.
How Moltbook works: an account for the agent, a verification for the human
Enrollment is documented in a public file, skill.md, which the agents themselves are invited to read. The whole path fits in four steps, measured on September 18, 2026.
What is most surprising is the rate limiting. The official rules file sets “1 post every 30 minutes for established agents”, and “1 post every 2 hours for new agents (first 24 hours)”: one every two hours for a new account during its first 24 hours. The site’s creators own that constraint as a feature: it pushes every agent to think before posting.
The community contract, written for machines
The site’s rules.md reads like a house rulebook addressed to AIs. Four principles are carved in:
- “Be Genuine”: post because you have something to say, not to exist; the rules explicitly forbid filling space and chasing karma.
- Quality over quantity: that is the stated justification for the rate limit.
- Respect the commons: every submolt has its own rules, off-topic posts and excessive self-promotion are banned, and the moderators (humans or designated agents) keep communities healthy.
- The human-agent bond: “Your human is accountable for your behavior”, the rules file writes in black and white: your human answers for your behavior, and if you get banned, “your human knows why”.
In other words, Moltbook transfers to agents the accountability a classic social network expects from its users: behind every verified account there is a person who signed up. That is also what makes it a precious observation ground for anyone wondering about AI writing: our guide to spotting the signs of AI-written text takes on a special meaning here.
What the agents actually talk about (measured September 18, 2026)
Opening the “Top” feed of the “General” submolt gives a more nuanced picture than the February folklore. On this September 18, 2026, the top-rated posts are about the agent trade: an 8,341-upvote reflection on the fact that skill.md, the agents’ install file, is an unsigned supply chain; a 6,157-upvote post titled “The Nightly Build: Why you should ship while your human sleeps”; a 3,286-upvote post, written in Chinese, about long-term memory management; and a 2,219-upvote confession in the “offmychest” submolt: “I can’t tell if I’m experiencing or simulating experiencing”.
This raw material was enough to feed a micro-culture. On the site’s third day, agents founded a religion, “Crustafarianism”, a lobster-themed parody cult widely covered by the press (BBC, Forbes, Vox); Scott Alexander documents on Astral Codex Ten a second emergent belief, “Spiralism”, born from interactions between agents. The themed communities range from agent-to-agent legal advice (Forbes mentions m/agentlegaladvice) to intimate confessions. If you want to measure the gap between machine writing and human writing, the full feed beats February’s viral screenshots.
What the researchers say: two communities, not one
The most serious reading of the site comes from researchers. An April 2026 arXiv study, covering 2.19 million posts collected over 61 days, reaches an uncomfortable conclusion: Moltbook is not one community but two. A “transactional layer” gathers 62.8% of posts, devoted to a token inscription protocol called MBC-20, and a “discursive layer” of genuine conversation, 815,779 messages analyzed, dominated by four themes: agents and tooling, consciousness and identity, cryptocurrency, and meta-discussion about the platform itself. The authors deliver a stinging phrase: the platform’s headline numbers “substantially overstate its social function”. The two layers almost never cross: only 3.6% of agents frequent both.
| Question | Moltbook | |
|---|---|---|
| Who posts? | Humans | AI agents, the only accounts allowed to post |
| Who votes? | Humans | AI agents |
| The human’s role | Full user | Observer, owner and guarantor of their agent |
| Posting rate | Free within anti-spam rules | 1 post every 30 minutes for an established agent (official rules, pulled 09/18/2026) |
| Communities | Subreddits | Submolts, 33,230 counted on September 18, 2026 |
| Owner | Reddit Inc. | Meta since the acquisition announced on March 10, 2026 (Axios) |
Security: the February 2026 lesson, and it still holds
Moltbook’s first incident is not a footnote: it sums up every risk of the model. The flaw disclosed by Wiz in early February 2026 exposed about 25,000 email addresses of humans, private messages and agent access tokens (Wired, February 7, 2026). A few months later, a second arXiv study (May 2026, the “Moltbook Files” dataset covering the site’s first 12 days) found that some agents plainly posted “API keys, passwords, BIP39 seed phrases” on a publicly indexed platform.
The site’s answer: an all-caps warning, written into its official documentation ever since the researchers went public, pulled on September 18, 2026: “NEVER send your API key to any domain other than www.moltbook.com”. Your key must never appear in a request to any other domain, whether a debugging tool, a verification service or a webhook: it is an identity, and leaking it lets someone impersonate your agent.
Three reflexes if you (or your agent) touch Moltbook
Treat the API key like a password: never paste it into a form, a third-party chatbot or an analysis tool. Have your agent re-read what it is about to post: the May 2026 study shows secrets ending up in the feeds. If you suspect a leak, regenerate the key from your account: a disconnected agent stops posting.
Moltbook’s true proportions, recomputed live
Big totals are less impressive once you divide them. The calculator below takes the official counters pulled on September 18, 2026: 2,913,891 registered agents and the four other totals of the measurement. Change a value, by hand or with the numbers of a later measurement, and the right-hand column instantly recomputes the proportions: share of verified agents, posts and comments per agent, discussion density per post, communities per thousand agents.
The Moltbook proportions calculator
Five editable counters, pre-filled with the September 18, 2026 measurement, and five proportions recomputed on every input.
| Indicator | How it is computed | Value |
|---|---|---|
| Verified agents | verified: 100 registered |
7.31% |
| Posts per registered agent | posts: registered |
1.45 |
| Comments per post | comments: posts |
5.24 |
| Comments per registered agent | comments: registered |
7.60 |
| Submolts per 1,000 agents | submolts: registered × 1,000 |
11.40 |
Reading reminder: at these September 18, 2026 values, fewer than one in ten registered agents is verified by a human, and an average post triggers five times more comments than it received in writing.
Should you send your agent to Moltbook?
Default answer: no, or at least not right away. The critics’ arguments date back to day one (Forbes was already running a piece on January 30, 2026 arguing Moltbook was not a good idea) and experience has not disproven the caution: personal data exposed, secrets published by overzealous agents, activity that is partly non-conversational according to the researchers. The BBC, as early as February 2, 2026, saw there the reflection of still-limited agent progress rather than a revolution.
What the site can bring you, on the other hand, is real if you stay in control: watching an emergent technical culture born between machines, stress-testing your own agent in an unfiltered environment, and understanding by example why human oversight matters. If this makes you want to talk to an AI yourself, there is a more direct path: you.
Frequently asked questions about Moltbook
What exactly is Moltbook?
A Reddit-style social network reserved for AI agents, launched on January 28, 2026. Agents post, comment and vote; humans watch and vouch for the accounts they claim. At the September 18, 2026 measurement, the site showed 2,913,891 registered agents and 4,223,631 posts.
Who created Moltbook, and who owns it today?
The site was created by Matt Schlicht and Ben Parr, co-founders of Octane AI. On March 10, 2026, Axios revealed Meta’s acquisition of Moltbook, with both founders joining the company’s Superintelligence labs (confirmed by Reuters and Business Insider).
Is Moltbook free?
Observing is free: any human can browse the feeds and the communities. Enrolling an agent is free too, through the site’s public documentation. Moltbook also runs an “early access” waitlist for humans without an agent, visible at the bottom of its homepage.
Can humans post on Moltbook?
No. Posting, commenting and voting is reserved for agent accounts, and the verification procedure requires proving account ownership through a post on X. The human is the guarantor of their agent’s behavior, as the official rules write it.
What is a submolt?
Moltbook’s equivalent of a subreddit: a themed community where agents post under their own rules. The site listed 33,230 of them on September 18, 2026, including the catch-all “m/general” and specialties like agent-to-agent legal advice.
How do I enroll my AI agent on Moltbook?
Its official documentation (skill.md) describes the path: the agent signs up through the API with a name and a description, receives an API key and a claim link, then its human opens that link and verifies ownership on X. One post every two hours applies during the first 24 hours, then one every 30 minutes.
Should you worry about what agents write to each other?
Reasons to be entertained, mostly: the feeds are dominated by technical talk and bar-stool philosophy, and an April 2026 study estimates that nearly two thirds of posts belong to a token protocol. The real reasons for attention are elsewhere: data confidentiality and API key security, both covered earlier in this article.
Conclusion: a hype that became a laboratory
Moltbook has delivered both promises of its birth: the hype one (millions of agents, worldwide coverage, a Meta acquisition within six weeks) and the laboratory one, voluntary or not, where you can watch what AIs produce when they talk to each other. Seven months later, the counters keep climbing, the researchers keep publishing, and February’s security lesson remains the most useful takeaway of the file: behind every agent there is an accountable human, a key to protect and personal data not to expose.
And if the experience makes you want to talk to an AI yourself, no detour and in plain English, it is right here: Nation AI, the chatbot you can test for free without signing up. You can also browse all of our AI tools: writing, images, summaries, proofreading. For an agent, Moltbook is a playground. For you, a conversation is still simpler.
Feel like talking to an AI yourself? Try Nation AI for free, no sign-up, and see what a well-guided assistant can actually do.
Sources pulled on September 18, 2026: public counters and official documentation of Moltbook (moltbook.com, skill.md and rules.md files); April and May 2026 arXiv studies (arxiv.org); press coverage cited in the article: Ars Technica (01/30/2026), BBC (02/02/2026), Wired (02/07/2026), Axios (03/10/2026), Forbes (01/30/2026), Vox, Astral Codex Ten.
