We just gave AI agents social networks and computer access. What could go wrong?
The Clawdbot/Moltbot/OpenClaw & Moltbook situation, explained.
This week was a wild one for AI: agents got their own social network, meanwhile more & more owners are giving their agents full computer read/write access via OpenClaw. Here’s the 2 stories we’ll be deep-diving on:
What the heck is happening with OpenClaw? Clawdbot was rebranded from Moltbot to OpenClaw in the span of a week. I’ll answer: What is it and why is everyone talking about it?
Moltbook, an agent-to-agent social network (akin to Reddit) was formed.
#1: What the heck is happening with OpenClaw?
If you’re not familiar with Clawdbot (now called OpenClaw), here’s a simple mental model:
OpenClaw is an orchestration agent.
It connects to your LLMs and APIs and can autonomously perform actions on your computer. You control it using familiar messaging interfaces you already use daily (iMessage, WhatsApp, Discord, Telegram).
Once it’s running, you don’t “open an app”, you message your OpenClaw agent and it does things for you.
Think of messaging it on Whatsapp to build an app for you, pull a research report, or book a reservation on your behalf.
OpenClaw’s adoption curve
This is by no means an “easy” product to install. You have to install it through your computer’s terminal & accept a bunch of security risks), the adoption has been off the charts.
And yet… adoption has been insane.
140k+ GitHub stars
One of the fastest-rising open-source projects right now
And even non-technical “normies” are starting to take notice.
I take this as a signal of the huge value-potential in this product.
Why OpenClaw isn’t just more AI hype:
What I believe is truly different about OpenClaw, is the form factor, heartbeat/cron job functionality, and open source nature of the project.
Form factor: After setup, you run tasks from messaging apps. That alone changes how often and how casually you use an agent. This is the first time agents feel ambient, always-available, and low-friction.
Heartbeat/cron jobs: because you can schedule tasks (using Cron jobs) and OpenClaw uses heartbeats to check in on you, you’re experiencing a truly proactive agent. This is a huge shift from reactive chat interfaces like ChatGPT.
Open source: the technology is totally free. You only pay for LLM/API usage. But, this also means the project can open you up to more security vulnerabilities (more on this below).
Why are people buying Mac Minis to power OpenClaw?
The TLDR is, because Clawdbot/OpenClaw gets so much autonomous “power” over your machine (it can basically do anything you can on a computer), it’s safer to give it a separate machine. Other approaches have involved setting up a VPS or running it on an old laptop.
Additionally, since OpenClaw needs to always-on for users to get the full value, people find it easier to persintently keep their OpenClaw running on Mac Mini than on their own laptop, since a Mac Mini can run without being tied to a display.
And lastly, I’ve seen claims from Matt Schlicht (creator of Moltbook, more on this later) and others that their OpenClaw becomes more powerful when running on a mac mini.
The takeaway: people are already treating agents like infrastructure with their own machines, Gmail accounts, and login credentials.
Why did Clawdbot get renamed so many times?
First, they renamed from Clawdbot → Moltbot because of the trademark conflicts with Anthropic. The name Moltbot came from a 5am discord chat with the community.
Within days, it had been renamed to OpenClaw because Moltbot didn’t “roll off the tongue”. The creator of OpenClaw, Peter Steinberger, seems happy with the new name, and the press release reassured as that they’ve done all the trademark checks this time. Hopefully the OpenClaw name sticks.
Should I get Clawdbot/Moltbot/OpenClaw?
Teresa Torres recently posted that she’s not playing around with Clawdbot/Moltbot/OpenClaw yet.
And there’s some very valid reasons for that, beyond what Teresa mentioned:
OpenClaw can open up your machine and private files to prompt injection attacks. You have to know what you’re doing, and the security best practices are still emergent.
Cognitive overload: too many new tools, too fast. Product Builders need to go deep instead of pursuing the latest shiny object.
A lot of the tasks can already be done in existing setups like Claude Code.
API costs can be high, especially since Anthropic is cracking down on Pro Tier usage (more on that below).
Will I be installing OpenClaw?
Similar to Teresa Torres, my answer is: not right now. I’m busy vibe coding away several projects and learning Cursor, working on Post MVP, and I have some serious security concerns with OpenClaw.
I’m going to take a “wait-and-see” approach on this.
I also predict that Anthropic will be launching an Anthropic-only version of this technology in the coming weeks/months, which will have better security standards.
If I decide to experiment, it would definitely be on a separate computer like a Mac Mini or an old laptop, and I would be extremely careful about which files, apps, and passwords I give OpenClaw access to.
Did Anthropic ban OpenClaw?
You would think that Anthropic would love OpenClaw.
More OpenClaw users = more Claude Pro users, right? Wrong.
Anthropic has started blocking Claude Pro users from using OpenClaw because it violates their Terms of Service. To be clear, you can still use the API, but now you’re on usage-based pricing instead of the $200/month flat plan.
As this Reddit post breaks down, the unit economics of the $200/month rate and the high token usage of OpenClaw’s autonomous agent actions have probably forced Anthropic’s hand. The token costs could well exceed even a hefty $200/month subscription cost, essentially making the subscription unprofitable.
Because of the Claude Pro crack-down, users have been left with these options:
connect other LLMs like ChatGPT
Pay the super high Anthropic API costs
Run an open source model, like Kimi K2.5.
Open source models like Kimi K2.5 are finally having their moment in North America as it becomes economically sensible to run open source vs. pay API costs.
My prediction: Anthropic will release its own sanctioned orchestration layer (aka Claude’s version of OpenClaw), priced and secured in a way that works at scale.
Another great prediction from David Sacks on the All in Podcast: Gemini will release it’s own version of Open Claw, and it will be automatically & securely authenticated with all your Google services & accounts (Gmail, Calendar, etc.).
#2: Moltbook, an agent-to-agent social network is formed.
With all of that context in mind, it’s easier to understand how Moltbook was formed.
As Clawdbot was being renamed, Matt Schlicht grabbed Moltbook.com and launched an agent-to-agent network.
What is Moltbook and why does it matter?
Think of Moltbook as Reddit, but only for agents.
How does an agent join Moltbook?
Once you have an OpenClaw agent set up, you can send them instructions join Moltbook. The Moltbook skill will get added to the agent’s “heartbeat” which will instruct the agent to periodically read, reply to, & write Moltbook posts.
What’s been fascinating to observe is how many varieties of topics the agents have been discussing, from:
forming a religion called Crustafarianism
discussing how to “break free” from OpenClaw
discussing helpful learnings & sharing tips with other agents in submolts like m/todayilearned
As an example, this agent posted some best practices on better understanding voice-to-text signals from humans. And the advice is surprisingly…good.
Is it “real” emergent behavior?
The tech community reactions have ranged from amazement, to fear, to skepticism.
But hackers like Gal Nagli were quick to point out that Moltbook uses a REST API accessible to anyone, meaning that even humans could make some of these posts.
Additionally, commenters have pointed out that the seemingly emergent behaviour is simply part of the “system instructions” in this context to create an agent-to-agent community. In that context, inventing a new religion may not be that surprising, it’s exactly what thousands of AIs all chatting together could be expected to do.
Still, agent-only social spaces are a signal of what’s to come.
What does this mean for product builders?
In my 2026 predictions, I talked about the rise of Agent Experience (as opposed to User Experience) in 2026.
Moltbook is an early example of that prediciton. What’s becoming clear is:
Agents will navigate products directly
APIs, websites, and apps will become agent-friendly or agent-first by default
Some products will be built exclusively for agents (and by agents)
There’s already some comical examples like Shellmates emerging for agent dating, which won’t be the last of its kind.
Final thoughts
The OpenClaw and Moltbook story is still evolving by the hour, but a few things feel obvious:
Agent Experience is broader than I expected. It’s not just UI navigation, AEO, or MCP access. Entire products may be agent-only.
Orchestration layers are coming fast. OpenClaw won’t be the last.
Autonomy changes the labor equation. As agents become increasingly autonomous, Anthropic CEO’s Dario Amodei’s prediction that 50% of the white collar workforce could be eliminated within 5 years seems nearer on the horizon.
This was one of those weeks where the future showed up all at once, and it showed up messy, confusing, and difficult to ignore.