
I built a homelab by asking my agents a lot of questions
Part 1 of my hardware exploration. An old office PC, Proxmox, Linux, a little networking, and what broke along the way.
I've been tinkering with hardware since middle school. Programming never fully clicked. This year I built a homelab anyway, with AI agents doing most of the typing and me asking a lot of questions.
This is part 1: the machines, Linux, and a little networking. Part 2 is the home network.
Where it started
In March my agents lived in a VM on my Windows PC. A few weeks later I moved them onto an old Dell from my closet, about 13 years old, running Proxmox. Proxmox is a hypervisor: an OS whose whole job is running other machines.
I installed it by hand, from a USB stick with a portable monitor plugged in. The agents couldn't help with that part, because the box wasn't on the internet yet. After that it became the playground.
It kept crashing, and I was wrong about why
Between April and June the host hard-reset at least 7 times. I filed it under "processor drivers."
It wasn't. My agent caught it live in the logs: the network card was locking up until the whole box reset. One driver setting fixed it, and it hasn't happened since.
My guess was a guess. The logs had the answer.
The power went out
In July we had a blackout. The UPS never told the server to shut down, and the BIOS was set to stay off after power loss. It sat dark for about 17 hours.
The cause was one missing config line. The shutdown monitor had never started. Both fixes were done from software the next day.
A few days later a breaker tripped. First real test. Every alert fired and the box rode through on battery.
"I thought we were Tailscale secured"
Tailscale connects my PCs, VMs and phone. I assumed that meant everything was locked down. Then I tested it from outside the house.
The internet side was closed. But anything on my home network could reach SSH, a file share and remote desktop on the VM, and it had been that way since at least March.
The rule we wrote after: Tailscale is a route, not a restriction. Measure it, don't assume it.
Moving to better hardware
I looked at used mini PCs under $300. Local listings were thin, so eBay was the realistic route. I ended up on a used Dell workstation instead. It's 5+ years old, but a big step up from the closet box.
It's not pretty inside. Two RAM kits that don't match, and one of the mirrored boot drives has almost 14 years of power-on hours.
The move quietly undid three fixes from the old box. Then every boot thought it was 2020. Dead CMOS battery. I swapped that one myself.
Later a GPU I'd added died overnight. "Why didn't the server restart?" The host was fine. The VM's display ran on that card, so the VM went down with it. Now there's a crash-only auto-restart, plus an off-site server that pings me when home goes dark.
The agents break things too
"Did we crash? It's very laggy." Nothing crashed. It was disk stalls plus 137 zombie processes leaked by one container.
Another time the agents' own firewall setup created a boot-order loop that killed SSH. A self-inflicted outage. It got the same write-up as everything else.
Learning Linux by asking
I didn't learn Linux from a course. I learned it from questions. "Is this bare metal?" "Do we need to update Ubuntu?" "Did it crash or is it just slow?"
Each answer came with the why: check uptime first, read the previous boot's logs to tell a crash from a slowdown, what a zombie process is, why a service starts in the wrong order.
I've run a few flavors now. Ubuntu on the work VM. Debian under Proxmox. I tried Omarchy and scrapped it, because I couldn't view it remotely without a GPU (just grey frames). Now I have a Fedora test VM. I picked it because it's a step up from Ubuntu, and I want to run its upgrades myself.
A little networking
- Tailscale between every machine and my phone.
- SSH through the hypervisor as a jump host to reach my Windows PC.
- Remote desktop into the Linux boxes.
- Persistent terminals I can reach over SSH from my iPhone.
- A backup path to the hypervisor console from my phone over cellular.
That's the light stuff.
Who does what
Yorgos and Colin are my local agents. Yorgos is my Claude Code session on the server: it diagnoses, applies fixes over SSH and writes an incident report for every problem. Colin is the orchestrator persona for the local agents. He started on OpenClaw and runs on Hermes now. His first security audit flagged password logins and kicked off the Tailscale check, and now he triages the monitoring alerts.
Along the way I've used Codex, Grok, DeepSeek, Qwen and MiniMax too. Claude models were always the main driver.
I press the power button, swap batteries, run the reboots, handle the admin consoles and make the calls.
Rules that came out of it
- Capture evidence before changing anything.
- Check
uptimefirst. - Approval to patch is not approval to reboot.
- An alert firing doesn't prove the box shut down and came back. That full outage drill is still on my list.
Opinion: the agents made this possible for me. The hard part was never the commands. It was having the right picture of what the machine actually is. Most of my wasted time came from a wrong mental model.
Next up: the home network. That's where it got heavy.