I’ve been tinkering with a little project called GreenClaw and I figured it was about time I wrote it up properly, partly because I think it might be useful to a few of you, and partly because GitHub stars don’t just appear by magic apparently.
So what is it? GreenClaw is a self-hosted, always-on AI agent. Think of it as a permanently running assistant ready to get things done — research, local LLM queries, dev workflow stuff, whatever you’ve taught it to do. It was loosely inspired by OpenClaw but built with the assumption that most of us aren’t running racks of enterprise hardware in a temperature-controlled room.
The hardware angle
This is the bit I’m actually quite pleased about. GreenClaw is running on a Lenovo ThinkCentre M920q — a tiny little desktop that I picked up secondhand. It’s the kind of machine that gets recycled out of offices when companies upgrade, so they’re everywhere and they’re cheap. We’re talking an i5-8500T, 16GB of RAM once I shuffled some sticks around, running Arch Linux headless on an SSD. No GPU. Idles at around 11w so not big energy bills either.
The local LLM side of things is handled by Qwen running via llama.cpp. It’s not going to win any benchmarks but for the tasks GreenClaw handles — routing requests, quick answers, local inference — it’s absolutely fine. If it needs to do something heavier it hands off to the Claude API, but that’s optional. You could run this entire thing for free if you stuck to a local model only. The Telegram bot is free. The hardware is peanuts. The Arch install costs you nothing but time and mild existential stress during setup.
What does it actually do?
Right now GreenClaw handles things like:
Answering questions via Telegram using the local Qwen model
Handing off heavier requests to Claude Code [using a pro account that has a set monthly bill]
Running custom “skills” — markdown files that define how it should behave for specific tasks
Sitting there patiently in the background 24/7, waiting for you to need it
The skills architecture is the clever bit (if I do say so myself). You drop a markdown file into the skills folder describing what you want it to do, and it picks it up. No code changes needed. I’ve been adding things like GitHub issue creation, research lookups, and blog posting — basically anything I’d normally have to go and do manually.
Why bother?
Honestly? Because I like the idea of having something running at home that’s mine. No subscription, no terms of service changing overnight, no data being sent somewhere I didn’t agree to. And there’s something satisfying about talking to a little box sitting in the corner of the room and having it actually do useful things.
It also scratches the tinkering itch nicely. Every time I add a new skill it feels like teaching it a new trick. Which is probably a bit sad but I am an Arch user so there we go….
If you fancy having a look
The project is on GitHub — search for GreenClaw or follow the link below. It’s a work in progress, the documentation is honest rather than polished, and I’d love a star if you think it’s interesting. Stars are basically the currency of validation in open source land and I am not above asking for them directly.
If you do try it out and hit any snags, raise an issue. I actually read them.
Any questions, feedback welcome…
Stay safe.