AI & Agents
I Run My Personal AI OS on a MacBook With a Broken Screen
The keyboard is dead and the display is shattered, and this MacBook is still my hardest-working machine. Here is how I turned it into an always-on home server for AI agents.
always-on AI home serverold MacBook as serverpersonal AI OS
There is a MacBook in my living room with a shattered screen, a dead keyboard and a trackpad that does nothing. It is plugged into the wall, online, and running agents all day. It is the hardest-working computer I own.
I wrote the original version of this in April 2026. This one is the setup guide I wish I had read before I started.
Why use a broken Mac as an AI server?
Because the hardware that is useless to you as a laptop is perfectly good as a box in the corner.
My Mac’s screen cracked and the keyboard died. I did not bin it. First I plugged it into the TV over HDMI and used it as a streaming box, which worked well enough. Then I noticed something uncomfortable. I was paying for VPS hosting to keep agents running around the clock, and I already owned a machine that was always plugged in, always online and doing close to nothing.
So I installed my agent stack on it and cancelled the VPS subscriptions. What I got was a home server with Apple silicon performance, local storage and no extra hosting bill.
What an always-on AI box actually does for me
“I built an AI setup” tells you nothing, so here is what runs on mine.
Overnight cron jobs. Agents run on a schedule while I am away from the machine. They search, cross-reference, research and write up structured briefs. A Telegram message with the summary is waiting for me in the morning.
A Telegram bot as the remote control. I made the bot through BotFather and connected it to my agent layer. When I am out, I message it and the home Mac answers. If I send a screenshot, the agents read it, pull out the structured data, add it to the right list and file it in the matching project folder.
A memory layer. Every conversation and piece of research goes through a local Ollama instance running several small models, each tuned for one job: summarising, entity extraction, routing and classification. The outputs land in ChromaDB, a local vector database. That is why a new Claude session does not need me to paste a giant context file. The relevant memory comes up on its own, including which project I am on and what I already decided.
A second Mac syncs the ChromaDB data and project state over rsync, so the context follows me between machines.
If you want the longer explanation of why those three pieces add up to more than a script runner, I wrote about it in what an AIOS actually is. I have also been packaging the memory and agent ideas into Holt, which is open source under MIT.
How to set up old hardware as an always-on AI box
This is the order I would follow. It is what I did, written as steps.
- Pick a machine that is already paid for. Mine was a MacBook I owned anyway. A broken screen does not matter if you only talk to the machine through software.
- Keep it plugged in and on your home network. The whole point is that it is always reachable.
- Install the agent stack and a local model runtime. I use Ollama for small, task-specific models.
- Add a local vector store. I use ChromaDB. Pick whatever you can run on the same box.
- Connect a Telegram bot. Create it with BotFather and point it at your agent layer. It is the easiest way I have found to reach a home machine from a phone.
- Schedule the overnight jobs with cron. Start with one task and add more after the first one has run cleanly for a few days.
- Sync to your working machine. I use rsync so both Macs share the same context.
I should say what I do not know. I have not benchmarked this against a rented server, so I cannot give you a speed comparison, and I would not trust anyone who says an old laptop always beats a VPS. It beat the VPS for my use.
Home server or $20 VPS?
People ask why I did not just rent a small server. These were my reasons.
| Factor | Why the home box won for me |
|---|---|
| Privacy | My memory layer, ChromaDB and project data stay on my home network unless I push them somewhere. |
| Performance | Apple silicon runs local models well. My Ollama models respond fast, with no cold start and no shared compute. |
| Cost | The Mac was already bought and the electricity is minor. The only recurring cost is my Claude subscription, which I would pay anyway. |
| Integration | It can reach local files, sync with my working Mac and use macOS-native tooling, which a remote server struggles with. |
A VPS still makes sense if your home internet is unreliable or you need a public address. Neither applied to me.
What changed in how I use AI
The surprise was that building this changed how I use AI, and not only where it runs. Once the agents had persistent memory I stopped treating every chat as a fresh start. I stopped re-explaining context and copy-pasting project briefs. That freed me to go straight to the thinking.
The broken screen turned out to be a feature of sorts. Since I could only reach the Mac through software, I had to build proper interfaces to it, which is how you should treat a server anyway.
Next on my list is better summarisation models, smarter routing and agent-to-agent handoffs, so specialised agents can pass work between each other without me in the middle. I think that last one is the hard part, and I have not solved it.
Checklist before you start
- Is the machine already yours, so the upside is a cancelled subscription?
- Can it stay plugged in and online all day?
- Do you have one overnight task worth automating first?
- Do you know how you will reach it when you are away?