The home lab
The lab
A Mac Studio in my apartment runs a small fleet of AI agents around the clock. It reports to a chat server I host myself, and I can check in from anywhere. This page shows how it fits together.
Lines pulse where data moves. Every box is real hardware or software I run or rent.
House rules
One writes, another checks
No agent merges its own work. A second reviewer reads every change, and pull requests to shared projects go to a person.
Everything has an off switch
Every scheduled loop has a kill switch and a spending cap, set per job before it ever runs.
One task, one worktree
Each job gets its own git worktree, so parallel agents never edit the same files.
Check the output, not the exit code
A job that “succeeded” has to show the thing it was meant to produce.
Write down mistakes
Incidents and decisions go into the vault so the next run starts smarter.
Local first
Small models run on the Mac Studio. Rented GPUs are for the heavy runs, with a budget per experiment.
Spec sheet
- Compute
- Mac Studio, Apple silicon. RunPod for burst GPU work.
- Agents
- Claude Code sessions, Hermes, llama.cpp, MLX, Ollama.
- Scheduling
- launchd jobs with full paths, minimal PATH and capped budgets.
- Comms
- A self-hosted Matrix homeserver. One room per concern, read from any device.
- Memory
- An Obsidian vault for decisions, mistakes and research notes.
- Network
- WireGuard for remote access; nothing on the lab is exposed directly.
- Code
- git worktrees per task; pull requests for anything shared.
Not on this page
What the agents research, trade or build next stays private. The house is public; the playbook isn't.
The reliability lessons are public, though. Read Six months of agents in my apartment, or see the Agent Ops Starter Kit.