Episode 1 – The Proxmox Marathon

"How to Build a Home Lab in 200+ Messages Per Chat"

I've been using ChatGPT for a while now — for quite a lot of things, if I'm being honest. Besides gaming and IT as a hobby, I've been an amateur photographer for years, and I upload my vacation photos to stock agencies. The images sit there waiting for interested buyers to download them, and whenever someone uses them in web or print media, the photographer gets their cut from the agency. Sounds chill. It's not entirely.

For the photos to reach any kind of usable quality, they need to be edited and the files tagged with titles, descriptions, and keywords. Up until recently, that was a pretty tedious process — the kind of work where you eventually start resenting your own vacation photos. ChatGPT then built me a Python script that automates exactly that. And when it comes to the editing itself, AI helped me level up my techniques and actually simplify them in ways I hadn't expected.

On the side, I'd also gotten my Linux system and ESO (The Elder Scrolls Online) tuned to a point where I was happy with both — every security hole and performance issue squashed. Long story short: I'd gotten a taste for this stuff.

The Idea Takes Shape

That's when an idea started brewing. With such a capable sidekick in my corner, I could finally make my dream of a home server setup a reality. I watched a ton of YouTube videos about Linux and home labs, and the pieces slowly came together in my head — sometimes quickly, sometimes in completely the wrong order, but they came together. I wanted to build a home lab with Proxmox as the hypervisor and create my own personal cloud. I was simply done paying some random provider every month for the privilege.

Inspired by the YouTubers who'd built exactly this kind of setup, and fascinated by the whole Proxmox-plus-mini-PC concept, I asked ChatGPT specifically about the technical requirements for buying a mini PC. No sooner said than done: I ordered a used mini PC on eBay, plus more RAM and a bigger drive — both installed the moment they arrived. The kitchen-table tinkering session was officially open.

And off we went! With ChatGPT's help, I managed to set up the Proxmox server and create my first VM running Linux Debian. After extensive research into cloud solutions and some solid tips from ChatGPT, I settled on Nextcloud combined with Tailscale — a VPN tunnel that lets me access my cloud from any of my devices without exposing anything to the public internet. Security through invisibility, basically.

Then I realized: the internal storage wasn't going to cut it. So an external SATA drive joined the party, meant to serve as storage for Nextcloud and backups. And just like that, I had it: my own cloud, my own data, accessible and editable from anywhere on any of my devices. For a brief moment, I felt like a living room-sized data center.

Network Chuck and the AI Agent Rabbit Hole

At the same time, I was going deeper and deeper down the AI rabbit hole. I stumbled across Open Claw and the concept of AI agents — an agent system that can actually do things autonomously on your machine. Fascinating, but riddled with serious security holes that the developer had deliberately left open to maximize functionality. More of an experiment than something I'd trust with my private data. Great concept, but I'm not handing my keys to someone who's left the front door wide open on purpose.

And then came a pivotal moment: a video by Network Chuck.

What he showed in that video genuinely impressed me. He was communicating through a terminal with an AI called "Claude" — one that can handle tasks autonomously, but in a much more controlled way. Claude? Hold on, isn't that the AI that Open Claw developer Peter Steinberger couldn't stop raving about! I had to find out more, and from that point on, the topic of AI agents had a firm grip on me and wasn't letting go.

ChatGPT and Its Limits

Because ChatGPT had its downsides — and by now I'd met every single one of them. Especially when troubleshooting complicated technical problems — Linux, gaming, photo editing — you'd quickly end up in a loop: the same broken solutions served up over and over again, like listening to a very confident person who's completely run out of new arguments. On top of that, I constantly had to ask ChatGPT to pause after each step and wait to see whether I'd actually completed it successfully — instead, it would immediately fire off the next five steps, along with a detailed catalog of everything else it could potentially do for me. Very ambitious. Not always helpful.

This tool was more entertainment than productive assistant for me. I wanted targeted solutions, one step at a time, with a proper memory and skill system. So: time to give Claude a shot.

I haven't regretted that decision for a second. Starting out as a browser-based chat, it walked me through my technical tasks — similar to ChatGPT, but a much better fit for how I actually work. I took out a subscription. With Claude, I could integrate memory both for the project and for general communication. You can configure how it communicates and dial in certain behaviors — that wasn't just interesting, that was exactly what I'd been looking for.

The Proxmox Marathon

So I kept optimizing my home lab with Claude: setting up a backup strategy, hardening the systems, step by step — and this time my assistant actually waited until I was done before moving on. Then came the brilliant idea: when I'm on vacation, I need a system that can power my machines on and off to save energy. Claude suggested solving this with a small computer — a Raspberry Pi — and a smart Wi-Fi power strip. Said and ordered. I was waiting on hardware again. As one does.

And while I waited, a new idea started taking shape: I could run Claude with an agent system just like Network Chuck's setup, right here on one of my VMs...

Actually a pretty good idea.