Whenever I say that self-hosting AI is not economical, people hear that I am against self-hosting. I am not: I run Qwen3.8 27B at home and think open-weight models are great. This post compares it with GPT-6 Luna on the same benchmark, and explains why batching makes datacenters win.

MindRoom grew one small group at a time at IonQ, until a company-wide demo made hundreds of people sign up in a day. What surprised me is that the heaviest users include engineers who already live in their coding agents.

Two years ago, a friend and I backed each other up through a TrueNAS VM on an iSCSI zvol. Now that I run NixOS, I replaced that machinery with zfs-tenant: OpenZFS delegation and a quota keep him inside one dataset, and a small SSH forced command plus zfs zone make sure he sees nothing else of my pool. His keys never leave his house, and a VM test checks each of those claims.

Friends keep asking how my homelab works, so I wrote it all down, starting from zero. Four NixOS machines run about 150 containers behind one Traefik front door, and with WireGuard and my own Headscale server I reach my self-hosted services from anywhere over encrypted HTTPS connections, while strangers on the internet get nothing. The part I think is the coolest is the balance: everything is declarative and lives in git, but with as little machinery as possible.

TypeSafe’s Jev, a new fast and cheap classifier model that doesn’t require fine-tuning, was suddenly, literally everywhere I looked, so I tried it in MindRoom. I now use it for small decisions I would never have spent an LLM call on, like whether a “thanks” should interrupt an agent that is still working.

A small Docker recipe for running Diction’s streaming gateway against Qwen3-ASR through Agent CLI, plus a test from a bar in the Netherlands, an ocean away from the GPU.

I have run coding agents in YOLO mode since I started using them. Instructions in AGENTS.md did not stop them from force-pushing or merging PRs, so hooks do. As the models got better, I added an override the agent may only use after I explicitly approve an action.

After replacing Proxmox and TrueNAS with NixOS, my GPU machine was the last one still on btrfs, picked years ago because forum threads had me convinced that ZFS and NixOS were a bad combination. What finally pushed me to fix it: I run AI agents in YOLO mode all day, the stories about frontier models wiping home directories kept coming, and my restic backups needed an hour and a half just to scan a hundred million files. The audit’s first finding: the backups had been silently dead for four months. This is the story of auditing my backups, verifying the restore path end-to-end, and wiping my daily driver so that nine machines share the exact same ZFS setup.

TrueNAS was the last appliance OS in my homelab. The build-system change was the last straw, but the itch had been building for a while: too much UI state, unclear defaults, and virtualization churn around Incus. This is the story of how I gave an AI agent root-capable SSH access with strict instructions not to change anything, used it to inspect the live machine, recreated the config declaratively, and built a VM rehearsal to prove disko would only wipe the boot disk.

MindRoom is an open-source system I built that creates AI agents living in Matrix. Because Matrix bridges to Slack, Telegram, Discord, and more, your agents follow you everywhere—with persistent memory, multi-agent collaboration, and 100+ built-in tool integrations.