March 14, 2026
Why I stopped using cloud AI and started running agents on my own hardware
How I run nine AI agents on a Mac Mini in my parents' house to monitor servers, write briefings, and audit configs for about thirty dollars a month.
For most people AI means calling an API. Pay per token, hope the service stays up, accept whatever rate limits they impose. That works fine until you want an AI that runs your operations instead of answering questions about them.

I run nine AI agents on a Mac Mini in my parents' house. They watch my servers, write morning briefings, audit security configs, track crypto markets, and manage projects. They cost me maybe $30 a month in API tokens, and they never go down because some provider scheduled maintenance at 2 AM.
The problem with cloud-only AI
Here is what kept biting me when I ran everything through hosted platforms.
Rate limits show up at the worst time. My monitoring agent checks 66 Docker containers every 15 minutes. Hit a rate limit at 3 AM when a container is crash-looping, and your smart monitoring is a silent log entry instead of an alert.
Latency adds up. The security audit agent parses configs, compares them against baselines, and flags drift. Round-tripping every call through a cloud API costs seconds per check. Multiply that across hundreds of config files and a quick audit takes 40 minutes.
There is no persistent state. Cloud APIs are stateless, and my agents need memory. The morning briefing agent knows what I read yesterday and what changed overnight. The project manager agent tracks where every task left off. You can bolt memory onto a cloud API, but then you are maintaining the orchestration layer anyway, so you might as well own the whole thing.
And the cost is unpredictable. Token pricing is fine for a chatbot. It is brutal for agents that run 24/7. I burned through $200 in a week of early testing because I did not realize how many tokens a monitoring loop eats. Now my agents keep their context local and only call external APIs for the heavy inference.
What self-hosted actually looks like
The setup is not complicated. The Mac Mini M4 runs my own orchestration layer on top of Claude, managing the agent lifecycles. Each agent is a Docker container with a defined scope, a schedule, and a set of tools.
The Homelab Monitor agent watches my Docker containers and alerts me before things break. Disk, CPU, RAM, container health, the basics that prevent 3 AM emergencies.
The Morning Briefing agent pulls my calendar, email highlights, project updates, weather, and news into a single digest, so I am not opening seven apps before coffee.
The Security Audit agent scans my Docker configs, Traefik routes, and firewall rules. It knows what normal looks like and flags anything that moved.
Each agent runs between $19 and $49. One purchase, no subscription. You download it, point it at your Docker socket and your API key, and it runs.
Why a Mac Mini
People ask why I do not use a big server for this. The Mac Mini M4 draws about 10 watts idle and has enough RAM for the orchestration layer plus some local model inference for lightweight tasks. The heavy inference still goes out to API calls, but the agent logic, scheduling, memory, and tool execution all happen on the box.
For the local inference I am not running anything I trained from scratch. The models are fine-tunes and abliterations of open-weight bases, with a small router model that picks the tool and a larger conversational model that writes the answer.
My production server, a Minisforum MS-01 with 96GB of RAM, handles the actual workloads. Docker containers, databases, media servers, the whole stack, while the Mac Mini just coordinates.
Total draw for the always-on gear (Mac Mini, MS-01, NAS, and networking) averages around 400 watts. The two Dell R730XDs only come online when I am testing HomelabARR PE's NAS and storage features, and they sit powered off the rest of the time. My monthly electric bill for the whole lab is around $40.
The part nobody talks about
The real win of self-hosting is the reliability patterns you cannot get from a cloud service, not just cost or control.
My agents survive internet outages. They monitor local hardware, so they do not need the internet to tell me a disk is at 95%. Alerts hit my phone over the local network first, internet second.
My agents have no cold starts. Cloud functions spin up, load context, process, and shut down. Mine are always running, so the monitoring agent's 15-minute loop finishes in 8 seconds because it is already warm.
My agents share a local network with everything they manage. No firewall rules to punch through, no VPN to maintain, no API gateway to configure. The monitoring agent talks to the Docker socket directly. The security agent reads config files off disk, and local operations carry zero network latency.
Getting started
You do not need a rack full of servers to try this. A Raspberry Pi 5 or any old laptop with Docker can run a single agent. Start with the Homelab Monitor, because it is the most immediately useful.
The full catalog is at agents.imogenlabs.ai. Nine agents, plus a bundle if you want all of them, and the Imogen Playbook walks you from zero to a fully orchestrated stack.
I also open-sourced HomelabARR CE, a Docker container management tool with a large catalog of app templates. It is free, MIT licensed, and it is what I use to manage all the containers these agents run alongside. You can try the live demo in your browser or grab the source on GitHub and run it yourself.