Today I ordered a Mac Mini. The smallest model, 16 GB RAM, nothing special. Not for work. Not for video editing. But to run a lobster on it.

Sounds absurd? Welcome to Moltbot.

Moltbot homepage, built with Moltbot

What on earth is Moltbot?

Imagine you could talk to your computer like you talk to a colleague. Not through a special app, but via WhatsApp. Or Telegram. Or Discord. Wherever you happen to be.

You write "Clean up my inbox" and your computer does it. You ask "What's on today?" and get a briefing. You say "Book me the flight" and it happens.

That's Moltbot. An open-source project by Peter Steinberger that collected more than 60,000 GitHub stars within a few weeks.

The community calls it "Jarvis", "iPhone moment", "early AGI". One user simply writes: "It's running my company."

What makes it special: Moltbot doesn't live in the cloud. It runs on your computer, with full access to your system. Files, browser, terminal, emails, all within reach. And you control it through the chat apps you already use.

There are lots of integrations. Moltbot can be controlled via Telegram, WhatsApp, Slack or Discord and can access many applications on its own:

Here is a fictional example chat for staying up to date and triggering actions via Telegram:

Fictional example Telegram chat, made with fakedetail.com

By the way, until today Moltbot was called Clawd, but that was a bit too close to Claude.

Until today Moltbot was called Clawdbot

What went well

The success stories are wild.

The insurance lawyer: User @Hormold reports that his Moltbot "accidentally started a fight with Lemonade Insurance" because it misread a reply. The result? The insurer reopened the case instead of rejecting it. Sometimes stubbornness pays off.

The Sora video producer: @xMikeMickelson asked his bot to make a video with Sora, "a little edgy". Five minutes later the bot had sorted out API keys on its own, set up watermark removal and built a complete workflow.

The university assistant: @pranavkarthik__ wanted access to his courses and assignments. He asked the bot whether it could build a skill for that. The bot built it, and immediately started using it.

The Whoop integration: @sharoni_k needed five minutes for the setup. Then he asked "What do you need to see my Whoop data?" The bot explained, fetched the data and now delivers automatic health updates.

The Nokia website: @youbiak built a complete website. From a Nokia 3310. By phone call.

What these stories have in common: people don't ask "How do I do X?", they say "Do X" and the bot finds a way.

What went wrong

Of course things break too. An AI agent with system access is like an overeager intern with admin rights.

The insurance fight (the other side): Yes, @Hormold's bot got the case reopened. But it also wrote aggressive emails without being asked. Sometimes you may not want your AI assistant fighting for you.

The self-shutdown: @bangkokbuild tells how he told his bot via Telegram to shut down the PC. The bot did it, and shut itself down in the process. Works as requested, but maybe not as expected.

The token flood: Several users report using up their Claude Max limit within a single day. @jonahships_ then had his bot build a solution itself: a proxy to his Copilot subscription. Creative, but expensive until then.

The heartbeat surprise: Moltbot can reach out proactively, via cron job at set times. @HixVAC was surprised when his bot suddenly wrote to him on its own. Well meant, but creepy if you don't know the feature exists.

The security disaster: As Günter Born reports, within 24 hours of the release hundreds of Moltbot servers were found on the internet without any authentication. One user gave his bot unrestricted access to his Signal account. Afterwards the credentials were publicly available online. A patch exists by now, but anyone who doesn't update stays vulnerable.

The architecture behind Moltbot

Moltbot isn't a simple chatbot. It's gateway infrastructure.

Chat apps (WhatsApp/Telegram/Discord/...)
    ↓
Gateway (runs locally on port 18789)
    ├─ Agent (Claude/GPT/local models)
    ├─ CLI (moltbot …)
    ├─ Browser control (Chrome/Chromium)
    ├─ Skills (extensible, self-generating)
    └─ Nodes (iOS/Android/macOS for camera, voice, etc.)

Installation is a one-liner (GitHub):

curl -fsSL https://molt.bot/install.sh | bash

Then moltbot onboard --install-daemon and the wizard walks you through it: model selection (Claude, GPT, local models), channel setup (WhatsApp via QR code, Telegram via bot token), workspace configuration.

Molt has great, extensive documentation:

Moltbot documentation

I'll share more details as soon as I can test Moltbot myself.

Why a Mac Mini?

Back to my order.

Moltbot needs a computer that keeps running. 24/7, ideally. Your laptop goes to sleep, loses its connection, and isn't reachable when you're on the road.

The community has solved this: Mac Minis as dedicated Moltbot hosts. The smallest model is enough, because the actual AI work happens in the cloud (Claude API, OpenAI API). The Mac Mini is just the orchestrator with Mac tools on board.

There are alternatives to the Mac Mini: Raspberry Pi, a Hetzner server, DigitalOcean, an old laptop. But the Mac Mini has advantages: the macOS companion app with voice wake, the iOS integration for camera access, native performance for browser automation.

I'll let you know next week how it's going.

Who is this for?

As of today: nerds. Installation is a one-liner, but you should know what a CLI is. The configuration is documented, but JSON files shouldn't scare you off.

But that's changing. The macOS app is already in beta. The onboarding experience is getting better. And the community is building skills that work out of the box more and more.

My prediction: within a few months someone will build a "Moltbot for normal people" version. One-click install, preconfigured skills, a friendly UI. And then it gets interesting.

Conclusion

When your AI assistant controls your computer, writes emails, fights with insurers, manages your calendar and analyzes your health data, where is the line?

Moltbot is a tool. A very powerful one. As @nateliason writes:

Hired my first employee today

Moltbot is what many people have been longing for. A local AI assistant that helps in everyday life or at work. BUT with great power comes great responsibility. I'm curious how the balancing act between security and capability will work out.

Moltbot isn't perfect. It's sometimes too enthusiastic, eats tokens like a hungry teenager and can break things if you're not careful.

But it shows where personal AI is heading: away from apps, towards infrastructure. Away from "ask me something" towards "do this for me". Away from cloud-only towards local-first.

Is this the future? I don't know. But I ordered a Mac Mini to find out.

The lobster is waiting.

⚠️ Important: If you now feel like trying Moltbot yourself: treat it like a new employee. That means: create separate accounts, no admin rights on sensitive systems, and watch what it gets up to at first. Moltbot is powerful, but power without control rarely ends well. Install updates, secure the gateway, read the documentation.


Moltbot is open source and free. You do need a Claude or OpenAI subscription for the AI backends, though. More info: molt.bot and the extensive docs.molt.bot. Also worth reading: Federico Viticci's article on MacStories and the StarryHope article on the Mac Mini trend.