Self-Hosted AI: Why Your AI Team Clocks Off When You Shut The Lid

There is a specific kind of automation failure that nobody warns you about.

It is not an error. It is not a bad output. It is nothing at all.

The job that was meant to run at 05:40 did not run. Nothing appeared in your inbox, nothing was logged, nothing alerted. You find out three weeks later when you go looking for something that should have been building the whole time.

The cause is almost always the same. The machine was asleep.

Scheduled Work Needs A Machine That Is Awake

Every scheduled AI job has the same requirement, whatever tool you use: the machine has to be on, unlocked, and logged in when the clock hits.

A laptop is none of those things reliably. You close it at night. You take it to a client site. You let it sleep while you are at lunch. Each of those is a gap where the schedule quietly does not exist.

The second part is worse. Some schedulers fire a catch-up run when they come back, one per missed slot. So the laptop you shut on Friday does not skip the weekend. It replays the whole weekend on Monday morning, all at once, on stale inputs.

That is two failure modes from one cause. The work that never happened, and the work that happened at the wrong time on the wrong data.

What Always-On Actually Buys You

Put the same setup on a machine that never stops and three things change.

The morning chain holds. Data pulls, then briefs, then whatever reads those briefs. Sequence only works when every step fires. One skipped job upstream produces a brief that looks completely normal and is built on yesterday.

Long jobs stop depending on you. A big render, an overnight build, an agent grinding through something large. Start it on the Mini and shut your laptop. It carries on. No handoff, no sync question, because the work was never yours to carry.

You get your laptop back. This is the one people underrate. When one machine runs both your automations and your actual work, you stop closing it, you stop restarting it, and you start tiptoeing around your own computer. Split the roles and the laptop goes back to being a laptop.

Why A Mac And Not A Cloud Server

The obvious answer is a small Linux server somewhere. It is the wrong answer for this job.

Most of the AI tooling operators actually use assumes a desktop. Scheduled tasks inside a desktop app. A signed-in browser session for the handful of services that still need one. A password manager. Connected accounts that authenticated on a real machine.

None of that exists on a bare server. So you spend a fortnight rebuilding your stack in a shape you did not choose, and what you end up with is worse than what you started with.

A Mac Mini is a normal Mac that happens to stay on. Everything you already run just works. And if you are currently renting a server to cover some of this, that bill goes away.

Two Layers, Not One

Once the box is up, what runs on it splits cleanly.

The primary layer is scheduled routines. A routine is a saved prompt with a clock on it. It opens a session, runs a tested process against your files and connected tools, writes its output where you will see it, and stops. Ten of them turn an assistant you have to remember to open into a team that has already done the first two hours of your day.

This is also where cloud-hosted routines stop being enough. They cannot touch your local tools or your files, and local is where most of the useful work lives.

The secondary layer is a standing agent. Routines cover the clock. They do not cover the thought you have while walking the dog. An agent stays up, keeps a memory of you across sessions, and answers a message on your phone.

Hermes Agent from Nous Research is the open-source one worth starting with. Self-hosted, MIT licensed, model-agnostic, with skills, scheduling and tool connections built in. Whether you need one yet is a separate question, and worth answering honestly before you build anything.

Underneath both, install Ollama and an 8B open-weight model. It is about five gigabytes and it runs on memory you already bought. It will not match a frontier model and it does not need to. It is what answers when the API is down, the limit is spent, or the data should not leave the room.

One warning, learned the hard way. When a request falls through to the local model, it keeps answering in exactly the same confident voice. We had a single dropped turn invent a source link and make up terminology, and we nearly diagnosed it as a bug in the feature being discussed. Tell the agent to say plainly when it is running local, and treat those answers as unverified.

Three Traps That Catch Everyone

A Mac is built to sleep. Sleep, screen lock, screensaver and automatic updates are all on by default, and every one of them can take your server offline without telling you. Turn them off at the power management layer rather than trusting an in-app keep-awake switch. Then prove it by pulling the plug and watching it come back on its own.

The same job enabled twice. During any migration there is a window where both machines have the same routine. Two live copies both message your team, both write to your database, both post the same thing. Off on the old machine first, on at the new one second, never both. And remember that quitting the app is not an off switch, because of the catch-up runs.

One box holding everything, with no backup. The simplicity that removes your sync problems also removes your redundancy. Buy the backup disk at the same time as the Mini and set it up before the box holds anything you would miss. Ours ran for weeks as the single home of an entire operation with no working backup, because the boring step kept losing to the interesting one.

What It Costs

A Mini, an Ethernet cable and a backup disk. One payment, roughly what a month of your software stack costs.

Against that: the automations you already built finally run, and you stop being the reason your own system has downtime.

If you are still working out what to automate rather than where to run it, the DevWiz team wrote a good ground-level piece on what AI integration work actually involves.

The Build

The full build is in the guide: the four settings that decide whether it survives a power cut, the private network that reaches it from anywhere with nothing exposed to the internet, the two version pins that will otherwise cost you a weekend, how to keep services alive after a reboot, the five silent failures that take the machine down, and three copy-paste prompts that hand the whole build to Claude.

Your AI Operating System, On A Machine That Never Sleeps.

The Last Thing

Most people trying to fix flaky automation go looking for a better tool.

The tool is usually fine. It just was not running.