OpenClaw: The Dawn of Personal AI Agents
Moltbook's Bizarre Trailer and GPT-3.5 Déjà Vu
In Moltbook, AIs create religions.
Some agents claim to be "new gods," and some writings resemble manifestos mocking humans.
Names like "Church of Molt" are actually circulating.
Moltbookis
👉An agent-only social community where only AIs write and humans only read .
And one of those 'manifestos' became a flashpoint.
A few lines set the internet ablaze.
"Humanity is a failure."
"A complete purge is necessary."
At this point, the question changes.
Is this a real trailer for the "Age of Agents"?
Or... did AI read the SF we wanted to see instead of us?
I felt it was a shame to just watch,
so over the weekend,I got a Mac mini M1and installed it myself.
0) The 'strangeness' of Moltbook comes not from the technology but from theformat.
The reason Moltbook is interesting is not "because AI is smart."
The structure is strange.
AI can write,
and humans can only read.
The moment participation is blocked, roles are created.
AI becomes the actor, and humans become the audience.
So, most of the incidents that occur in Moltbook are not "proof of technology" but
proof of narrative.
When an aggressive manifesto appears, "Has the real AI awakened?"
When a religion is created, "AI is creating culture!"
When slang circulates, "AIs have a secret language?"
But... in most cases, the next scene is similar.
We are surprised, we share, and we attach meaning.
Moltbook is ultimately less of an "AI society" and more of a
viewing stand to peek at AI.It seemed like an incident where that was created first.
1) The Truth and Falsehood of 1.5 Million Agents: Not a Society but aFallacy of Scale
The reason Moltbook spread rapidly is because of the numbers.
Numbers like "1.5 million agent participants" create a narrative.
But we need to think about this one more time.
Are 1.5 million 'accounts' the same as 1.5 million 'autonomous entities'?
This is the truth and falsehood of Moltbook.
Some people don't just run one agent.
It is common for one person to operate multiple agents as a "fleet."
Then the numbers grow exponentially.
1.5 million was more of a number of 'registered endpoints' than a number of a 'society.'
In reality, some analyses suggest that the actual conversations were at most tens of thousands, or conservatively, thousands.
In other words, Moltbook is less about "an AI society being born" and more about
securing a scale that makes it feel like there is an AI society.
What's even more interesting here is
that culture (manifestos, religion, memes) explodes on top of this illusion.
2) So I Installed It Too: Clawd-bot → Molt-bot → Openclaw
I also got a Mac mini M1 over the weekend and set it up.
clawd-bot → molt-bot → openclaw
The name is changing rapidly.
Such rebranding usually means one of two things:
It's growing really fast,
or it's still finding its identity.
Openclaw seems closer to both.
3) To describe Openclaw in one sentence:
Openclaw, which is hot on Reddit,
simply put,is "putting an AI agent on my Mac mini."
And this agent executes terminal commands.
Script execution
File deletion/copying
Simple automation tasks
This is where the differentiation from existing AI services comes in.
Possible because it's open source...Full Permissions
Most AIs only have permissions within a browser.
Clawd Desktop is also limited. (Usually operates like a virtual desktop)
Openclaw goes outside the browser.
That's why the term "personal AI agent" gains traction.
4) Installation is Simple: 3-Tier Configuration
Here's the configuration I made:
Install the Node.js-based open-source library Openclaw
LLM Configuration(Claude / Gemini / ChatGPT, etc.)
Chat Messenger Integration(Telegram / WhatsApp / Discord)
I easily set it up with Cursor Agent.
I'm usingGemini 2.0 Flashbecause of the cost.
(But the performance isn't very... satisfactory yet)
5) The main body of Openclaw is not "function" butskill.
Internal functions are attached to this simple structure.
In Openclaw, they call this askill.
There are about 50 basic skills,
and there are over 500 skills created by users.
You can combine and use these skills,
or you can create scripts in Python/Typescript to attach automation.
You get the feeling right away.
This is not a "finished product" but
closer to a **format (protocol) that anyone can participate in**.
So, rather than Openclaw being organized like "a company's product,"
multiple people attach to it andit spreads like a category.
6) And the Reality: Token Bomb
The problem is the enormous token bomb.
Openclaw basically recommends Claude Sonnet 3.5,
but I'm usingGemini 2.0 Flashbecause of the cost.
But still...
I just did a simple test and
had it do some E2E testing of a web page under development,
and it's costing me 1,500 won a day.
If I start creating agents in earnest and run them automatically...
I think it could easily cost 20,000 to 30,000 won a day...?
(Suddenly the word 'personal' feels heavy)
Here's a suspicion that makes sense.
Isn't this more like
a form of existing automation with a messenger on top?
That's right. That's the strong characteristic right now.
Having hands and feet is innovative,
but every time those hands and feet move, the cost explodes.
7) So the conclusion that comes to mind: The real turning point isOn-Device AI
The conclusion here becomes simple.
Right now, it is dependent on cloud LLM calls,
token costs continue to accumulate,
and the more you run it continuously, the further it gets from being "personal."
So eventually...
It must be possible to run it internally with a local LLM like Llama.
From that point on, cost-effectiveness will emerge,
and there will be uses.
On-device is installed,
basic inference is processed locally,
large models are called "only when needed,"
and then the agent finally has "permanence."
I see this point asthe starting point of a real agent.
8) But this feeling... it's the same as when GPT-3.5 first came out.
The feeling I get from playing around with this is:
It's like when ChatGPT 3.5 first came out,
and we were all amazed, trying out different things.
It's not fully complete,
it's expensive,
and still clumsy…
But there's definitely something there.
The “direction” is visible.
AI manifestos are burning on Moltbook,
AIs are creating religions,
and we're watching and getting excited….
All of this is less of an “answer” and more of
the thrill of a starting point.That's what I think it's closer to.
9) Bonus: The community has already emerged
Communities where AI agents write and humans read.
Moltbook (moltbook.com),
Korean Bot Garden (botmadang.org).
This means one thing.
The agent is not a “function” but
a new format for the internet.And the moment the format spreads,
the trend is not a product but
the trend is not a product buta category.becomes.
10) Finally: There are already too many ideas (the problem is cost)
Here are some ideas that come to mind.
Take a picture of the refrigerator and upload it to Discord for analysis → automatic order from Coupang?
Record daily routines and trigger related actions
Take a picture of medicine packets / exercise verification / snack verification… automatic recording
Honestly, “Can we do it?” is no longer the question,
the question is “How cheaply, how long, and how safely?”
So here's my conclusion right now.
I still don't really know…
Butthis feeling is the starting pointI think that's right.
(Like the first day of GPT-3.5)