Clawdbot/Moltbot - Beyond the Ugly Name

Antoine Gagné

— 5 min read
Clawdbot/Moltbot - Beyond the Ugly Name

Let's get it out of the way: I think Moltbot is a pivotal moment for Generative AI.

Let me elaborate. First, for anyone not familiar, Moltbot [1] is a personal AI assistant that you run on your own devices. Well, probably not all of it, the LLM part is probably still handled by one of the big players. But Moltbot orchestrates the LLM into a powerful assistant.

Wait. What does that mean? ChatGPT is already an assistant. Gemini, Claude are assistants. What does this assistant do that others don't?

What's special about Moltbot

Let me start by offering the Github star count chart:

Moltbot github stars as of Jan 28 2026

I'll let you try to guess when they were featured on Hacker News. Clearly, they do SOMETHING others don't, it arose a special kind of interest.

I've tinkered with it for 2 days straight, so I'm in an absolute expert position to pontificate on this. Here's what I think this does that's new, and why I think it's a pivotal moment:

  • You set it up, and it helps you connect it to Whatsapp, Telegram, Discord, etc... There's no friction in accessing it. It lives where you spend time.
  • It has FULL access to the computer it's setup on. Whether that's a good thing or not privacy/security-wise is a different question - it remains that this gives it unprecedented autonomy on how to perform tasks. It can click on stuff. It can download files. It can run scripts.
  • It has a Skills hub - Anyone can create and publish a tool. Literally - if you can explain what you want to do in human words, you can make a tool and make the AI use it. These tools may be as "simple" as text-based guidelines or provide scripts, API documentation, etc... that the assistant can make use of. It's modular advanced capabilities that you can provide to your assistant.
  • Agentic Behavior - You give agents a task, and it can use all the tools at its disposal to make progress towards the goal, iteratively. Moltbot might not just respond once when you message it. It'll send updates along the way, ask for clarifications or help if it gets stuck. Agents aren't new (by GenAI's standards) but it absolutely makes sense here.
  • Cron Jobs/Heartbeat - This one I'm an absolute fan of. You can setup prompts to "go off" at set times or regularly during the day. The assistant receives the prompt and can work on it. You give it instructions, in plain language, and it does what it needs to.
  • Files folder - Not sure what the name for this is, but Moltbot has a dedicated folder containing some files that gives the assistant guidance/information on a variety of things. Some examples:
    • IDENTITY.md: Explain to the Assistant "who they are". If you're into role-playing, that's the fun part. But besides that, you can give instructions about how to reply (detailed responses vs succinct ones, what persona to take on when replying, etc...)
    • USER.md: Explain to the assistant who you are. Are you someone technical or not, what's your field of expertise, etc... It can also contain likes/dislikes, family situation, etc... The assistant is someone you'll "work" with, they need to know a bit about you.
    • TOOLS.md: Give the assistant guidance on HOW to use the skills it has access to. As an example, if you use an image generation skill (which uses Gemini or OpenAI's image gen API), you might tell the model to not use the high-quality option every time - it gets expensive. This is not documentation about how the skill works, it's documentation about how to use the skill in YOUR setup.
    • MEMORY.md: Give the assistant a persistent memory between sessions. Self-explanatory.

All of this is pretty nice on its own, but the thing that ties this all together, in my opinion, is the fact that the assistant has FULL CONTROL over all of this. It can autonomously create new cron jobs, make edits to its memory, update a skill... In practice, it is a bit messy and clumsy, but this is clearly a first foray into self-improving systems. This is huge.

Not all of Moltbot's success can be pinned on its features though. This would never be possible if not for how good LLMs are these days. I haven't seen the prompt it sends to the LLM, but with how fast my Claude session usage % fills up, I can imagine there's a LOT in there. You need models that have a fantastic ability to "pay attention", so that it remembers to use that 2-sentences-long instruction (in 700+ pages of text) about using tool XYZ whenever you ask it to.

Some Examples

Let's say I want a daily brief. I send a message on WhatsApp to the assistant.

"I'd like a daily brief with a weather report for my city, an overview of the day's meetings and a list of new emails that would require urgent attention. Set it for 8am"

With this prompt, the assistant creates a cron for 8am to fetch the weather (it has a skill for this), fetch my calendar (it has a skill for this), to fetch my most recent emails (it has a skill for this) and send it back to me on WhatsApp. If I want to adjust this, I can just tell the assistant what I want to adjust.

You could ask it to text you to remind you of people's birthdays. You can have it monitor a topic (a web search every day) and summarize the news around it.

You can have it deliver hologram food... Not yet, but soon. Probably. (Blade runner JOI reference)

And where it becomes even more powerful: It is SO easy creating new skills. It took me about 1 hour to setup a chatterbox (Text to speech) model, wrap it in an API to serve, write the documentation on how to use it and give it to my assistant. And now, I can ask it to send me voice messages. I've used it to have the assistant research a topic and send me a summary I can listen to at my leisure and send a little pep talk for my 3pm daily motivation slump.

Closing Thoughts

There's two big caveat to all this at the moment.

1/ It wastes tokens SO fast. It loads up a LOT of context in each prompt, and agentic capabilities, computer use as well as tool use means it generates enormous amounts of text that you don't ever see.

2/ You still need to be fairly technical to set this up. You need to get API keys, set environment variables, edit JSON files, etc... Not something your typical grandma can do.

But this thing is like, 3 months old. It's the most rough and unpolished it'll ever be, and yet it's seeing this much attention. It is frictionless, customizable, autonomous and self-improving. I think this is what will make Gen AI "click" with the average population.

I'm calling it, among the other crazy development AI goes through this year, we'll start seeing actually good personal AI assistants. This will be a catalyst for mass adoption.

[1] By the way, you would have read this a couple days ago, it would have been named Clawdbot. Apparently they received some nice letters from people at Anthropic about the name, so had to find something new...

Read more