I started this for a few reasons. The first is control over what I run and how it runs. I’m sick of certain companies making wholesale decisions about things I use, and running them myself is how I get a say.

The other two reasons are closer to work. Building a platform end to end on my own hardware is a learning exercise that feeds straight back into my architecture work. And I wanted to see how far I could push Claude, and whether an LLM can really work autonomously.

Before this started, the house ran on two boxes. Home Assistant lived on a Raspberry Pi. Plex and the download apps ran in Container Manager on a Synology DS918+. None of it was code. It was all clickops, and if something changed, it changed because I clicked it.

The plan to replace that is dated 17 October 2026. The repo that built it had its first commit on 9 November. By 6 pm on the 14th there were 574 commits on main, nine VMs, a CI runner on a network of its own, and an agent that can take a GitHub issue through design, build and review in containers. Claude wrote almost all of it. This series is about how that went, and part one is the boring but necessary bit: what got built, and the decisions it rests on.

Three weeks on paper

For three weeks nothing ran. The plan was a markdown file dated 17 October, and the server, a Minisforum MS-03, was bought a week later. Over the next four days came every service with where it runs and how it’s deployed, the policy for which updates merge themselves, the VM names, and a clean-sheet network design.

The plan’s goals were one machine running every VM, the Pi retired, and the media devices off Wi-Fi, with local voice control in each room and a self-hosted assistant further out. Neither of those exists yet. The services doc added the other aim: replace paid services where a self-hosted one is good enough, so Immich for photos, Paperless for documents and Audiobookshelf for audiobooks.

The MS-03 arrived barebone, so RAM and NVMe were separate purchases. It has a Core Ultra 9 386H, 64 GB, two NVMe drives as a ZFS mirror, and two 10G SFP+ ports. Reviews put it at about 14 W idle, which matters for something that never turns off. My own measurement came out nearer 34 W, and the small case matters too, as part four will show.

The services document is the one that did the most work. It lists every app with who uses it, what it needs and where it runs, and its “Dropped” table is as useful as the rest. Pi-hole is out because NextDNS already filters at the gateway. Vaultwarden is out because 1Password covers it. Metabase is out because Grafana and Claude between them answer the same questions without a 2 GB Java process.

The shape of it

The naming rule went in on 27 October: every machine, physical or virtual, is named after a mythical creature whose myth hints at its job. The server is a Proxmox VE host called hydra, because it’s many heads on one body. The Synology became dragon, since it sits on the hoard. Proxmox’s description field on each VM holds the plain role as well, so nobody has to know their folklore to find Plex.

VMJobWhy the name
argusMonitoring and control: Grafana, Loki, VictoriaMetrics, Alertmanager, Uptime Kuma, ntfy, Komodo CoreThe hundred-eyed watchman
cerberusLogin only. Pocket ID, and no other appGuards the gate
cyclopsPlex, with the iGPU and NPU passed throughThe smith at the forge, where the heavy hardware lives
goblinThe download apps. No access to anything but the media shareGrabs loot, and isn’t fully trusted
golemPaperless, with the family-facing apps to followDoes what it’s told
brownieHome Assistant’s new home, built empty on the 14thThe household spirit that does the chores
djinnWhere Claude livesBound to serve, and grants requests within the rules
koboldCI, alone on a network that can only reach the internetWorks down the mines, out of sight of the house

There’s one more VM that I’m leaving out of this series. A public edge VM called sphinx is designed but not built. Like its namesake at the gate, it won’t let anyone past who can’t answer, and the answer is a client certificate.

Every one of them is a full VM, with no LXC containers anywhere. Proxmox itself recommends a VM for Docker rather than a container, and keeping to one guest type means one Ansible baseline and one backup method. Each VM runs Ubuntu 26.04 rather than Debian because I wanted the newer kernel and Intel drivers for GPU and NPU passthrough on a chip this new. The app VMs each run their own Caddy for TLS, Grafana Alloy for logs and metrics, cAdvisor, and the Komodo Periphery agent.

cerberus is deliberately the smallest and dullest VM in the house. If it goes down, existing sessions carry on, admin access has a break-glass path, and Home Assistant and Plex don’t notice. That only holds while nothing else runs there, so the monitoring lives on argus instead.

How a change reaches a server

The repo is the interface. Everything that runs is described in a private GitHub repo, and the rule written into Claude’s instructions is that nothing the repo automates gets done by hand.

homelab/
  ansible/        # VM baseline: users, Docker, agents, reboot slots
  proxmox/        # vms.yaml + pve-vm.py, which creates and reconciles VMs
  komodo/         # Komodo servers and stacks as TOML
  dns/            # records.yaml, synced to the gateway
  stacks/<vm>/<stack>/compose.yaml   # plus secrets/*.tpl with op:// refs only

A new VM is an entry in vms.yaml and a run of pve-vm.py. A new app is a folder under stacks/, a compose file, and a [[stack]] entry in Komodo’s TOML. Secrets are templates holding 1Password op:// references, rendered on the server at deploy time, so the repo never contains one.

When a change merges, Komodo’s Resource Sync reads the TOML and only the stacks whose folders changed get redeployed. The result goes to ntfy on my phone. Renovate opens a PR for each new image, CI and a Claude review check it, and small updates may merge overnight. That hasn’t actually happened yet, and the review gets its own part later, because release notes are text from strangers and it’s an LLM reading them.

The decisions that held

Every choice in this section was my call, the naming theme included, and I’d defend all of them. A few, with what they beat:

ChoseOverBecause
One Komodo stack per appOne stack per VMAn update restarts one app, not a whole VM’s worth
KomodoPortainer, DockgeGit stacks and Resource Sync: rebuilding Komodo is pointing a sync at the repo
Periphery as a systemd servicePeriphery in a containerA containerised agent can’t redeploy itself mid-job
GitHubForgejoDeploys don’t depend on the homelab being up
NFS for VM mountsSMBMetadata speed, hardlinks for the download apps, no case-insensitivity surprises
Home Assistant ContainerHAOSRenovate updates it like everything else
No SwarmDocker SwarmOne physical host, and Pocket ID’s SQLite can’t run active-active anyway

The one that didn’t hold was remote access. The plan said WireGuard for everyone, and that lasted until 9 November. An always-on VPN on relatives’ phones that I don’t manage means captive portals break it, my outage becomes their outage, and they get far more reach into the network than a Plex request page needs. Tailscale was rejected over battery drain on iOS, Cloudflare Tunnel over its 100 MB upload limit, and a VPS reverse proxy because it puts the trust anchor on a machine I don’t own and costs US$6 a month for little gain. What replaced it is a single edge VM taking client certificates on one port, which is the design for sphinx. The design keeps WireGuard, which isn’t built yet, for my admin access and for one app that can’t present a certificate.

The week

The commits by day give a sense of the pace:

DayCommitsWhat landed
Mon 9 Nov69Repo restructured, first VMs from vms.yaml, monitoring on argus, Pocket ID, Plex and the download apps moved off the NAS
Tue 10 Nov90Komodo git stacks, PR validation, deploys from the runner, Renovate
Wed 11 Nov62djinn built and Claude given its own accounts, Paperless’s VM
Thu 12 Nov98NAS audit, nightly VM backups and a tested restore, Paperless live, usage telemetry
Fri 13 Nov141The out-of-memory loop, the throttling, CI minutes running out and the isolated runner
Sat 14 Nov (to 6 pm)114The backlog as issues, the story launcher, the first story run through it

Claude’s own GitHub account made 341 of those, 129 of them PR merges. Under my name there are 131 commits with an Assisted-by trailer, mostly from before Claude had an account of its own, plus 61 merges I clicked. Not one non-merge commit is mine alone. Renovate made 32, and the nine from “Story worker” are the launcher’s headless runs, which are the subject of part six.

Early on that took a lot of my time. I was working out how each piece worked, and going through it with Claude to find out whether each thing could be done by hand before anything automated it. Then my part tapered off. By the end of the week I was mostly involved in what a story or task should achieve, and the design and implementation were left to the agents.

It wasn’t tidy. On the Friday the monitoring VM got stuck in a loop of out-of-memory kills, because I didn’t understand how Proxmox ballooning decides who gives up RAM. The same day GitHub Actions minutes ran out and the CPU started throttling. Those get two parts of their own.

What’s coming

Part two is the working agreement that let Claude do this: its own accounts, full sudo instead of an allowlist, change classes, and the rule that it merges its own PRs. Part three is about guardrails written as prose and what they turned out to catch. After that come the incidents, the CI runner, and the launcher that runs agent workers without trusting a word they write.

Assisted by Claude Code (claude-opus-5-5). The experiments, bugs and conclusions are mine. How I work with AI tools is in The AI pair programmer.