Zero GPU, zero dollars: a Linux rookie's crew of free AIs

Rabbid Raccoon — Wed, 07 Oct 2026 20:09:12 +0000

A few weeks ago I couldn't have told you what sudo does.

Today my laptop runs a whole crew of AI agents. They babysit my backups, my home lab and my gadgets, and report back every day, in character. One of them is Rick Sanchez. One of them is a British butler who swears like George Carlin. None of them cost a dime, and the laptop doesn't even have a graphics card.

I'm not a programmer. I built all of this by describing what I wanted to Claude, testing it, complaining, and testing it again. This post is what I learned, for anyone else who's starting from zero.

The house rules

Everything I build follows four rules:

If it ain't free, skip it. Free and open source first, every time.
Local beats cloud. Private, self-hosted, works when the internet doesn't.
Prototype first. Test wild ideas on cheap, spare hardware before building them for real.
Nothing starts at boot unless I say so. My laptop, my rules.

Rule 1 sounds limiting. It turned out to be the most creative constraint I've ever had.

The setup: no graphics card

My main machine is a regular laptop with 15 GB of RAM and no GPU. Everyone online says you need a big graphics card for AI. You don't, as long as you pick the right tools:

Omarchy (Arch Linux + Hyprland) as the operating system.
Ollama to run small AI models on the CPU: Qwen3, Llama 3.2, Gemma 3, Mistral. The 3-8 billion parameter ones are the sweet spot - slow-ish, but they work.
A "brain switch" (LiteLLM) that sends requests to the free tiers of Groq, Gemini, Mistral, OpenRouter and Cloudflare when I'm online, and falls back to the local models when I'm not. One address, many brains. It even worked out which free service is fast for which job: my coding agent went from 1-3 minutes per step to about 4 seconds once I put Mistral's Codestral first.

I keep the full list of models I actually use in a Hugging Face collection.

The crew

The fun part. Each crew member is a scheduled job in n8n that gathers facts, asks an AI to write a short report in character, and drops it into my daily notes (sometimes out loud):

Rick gives the morning briefing (PG - kids use this computer).
Birdperson guards the backups.
Beth checks the home lab apps.
Morty reminds me about my Linux lessons, which is fitting.
Mr. Poopybutthole keeps my GitHub and Hugging Face looking good. Ooo-wee.

They all live on Mission Control, a Rick and Morty themed dashboard that also launches my apps and searches an offline library (Wikipedia, medical encyclopedias, repair guides) when the internet's out.

Joshua, the voice butler

My favorite thing I've built. Joshua listens all the time but stays silent until he hears his name. Then he answers out loud in a British accent with the manners of Jeeves and the mouth of George Carlin. Hearing is fully offline (Vosk), the voice is Kokoro, and the brain is a free model.

My lesson from Joshua: keep the AI honest. Early versions "opened apps" that they never actually opened. Now there's a list of real actions checked before the AI gets a say, so he can't pretend.

Sharing it

Everything is public, free and MIT licensed:

🧪 Try it right now: Ask Rick - a PG Rick whose AI brain runs inside your browser. No server, no account.
🦙 Run them yourself: ollama run 2db2du/rick or ollama run 2db2du/joshua (on Ollama) - no GPU needed.
🐙 The code: github.com/2db2du-byte
🤗 The whole setup + every model I run: Rabbid's Lab and the collection
🌐 Homepage: 2db2du-byte.github.io
🦋 Say hi: @rabbid-raccoon.bsky.social

Before anything goes public, a script builds cleaned copies of my projects and scans them for personal info (names, drive serials, home paths, keys). If it finds anything, it refuses to publish. Beginners: build that safety net first.

What I'd tell another beginner

You don't need a GPU to start. Small models on a CPU are slow but real.
Describe what you want in plain words. AI coding tools are good at turning "I want a butler that only talks when I say his name" into working code.
Test everything. "It should work" and "it works" are different sentences.
Make it fun. I learn more from a burping Rick than from any tutorial.

Next up: actually finishing Unit 1 of my Linux course. Morty keeps reminding me.

Zero GPU. Zero dollars. Zero excuses.

- Rabbid Raccoon 🦝