Mac mini specs for running AI agents 24/7

Most buying advice for "a Mac mini for AI" assumes you are running models on the machine. It tells you memory is everything, and to buy as much of it as you can afford.

That advice is correct, and it is aimed at somebody else.

If your agents call an API — Claude, Codex, or anything else over the network — the model never touches your memory. It runs in somebody else's data centre. Your machine is doing orchestration: holding context, running shell commands, moving files, waiting on HTTP. That is a light job.

I run three Mac minis this way, continuously. Here is what actually matters.

The machines

Three Mac minis, each an M4 Pro with 24GB of memory and a 512GB internal disk.

They run many independent agent tasks at once, driven by orchestration layers on top of Claude and Codex. Not one agent at a time. Many, in parallel, for as long as there is work.

In that time I have not hit a resource wall. Not memory, not CPU, not thermals. The machines are quiet enough that you would not know they were working.

That is the finding, and it is worth stating plainly because it contradicts almost everything written on this topic: 24GB was never the constraint.

Why memory stops mattering

When a model runs locally, the whole thing is loaded into memory and stays there. A 70B model at reasonable precision needs tens of gigabytes before you have done anything useful. That is where the "buy 64GB" advice comes from.

An API-driven agent holds none of that. What it holds is conversation context, a working directory, and whatever tools it has open. Measured against a resident model, that is nothing.

So the memory question inverts. Instead of "how much do I need for the model", it becomes "how many tasks do I want running at once". Those are very different numbers, and the second one is small.

Where I would spend the money instead

Storage, not memory. 256GB works. 512GB is better, and the difference is felt more than the memory difference is. Agents accumulate: checkouts, caches, logs, build artefacts, the debris of tasks that ran while you were asleep. Disk is the thing that fills.

Apple storage is expensive, and it cannot be upgraded later. That makes external storage the sensible move: a Thunderbolt or USB4 NVMe enclosure gives you room for a fraction of what Apple charges to go up a tier.

A second machine before a bigger one. Three separate minis running independent work has been better than one large machine would have been. Tasks are isolated, and one machine misbehaving does not take the others down.

The two things that actually break it

This is the part the spec sheets do not cover, and it is where all my real problems have come from.

Your internet connection

The agents are useless without it. A local model keeps working when the line drops; an API-driven agent stops mid-task. Everything you have built depends on a connection you do not control.

That reframes what "reliability" means for this setup. It is not a hardware question. It is a question about your router, your line, and whether the machine is on Wi-Fi or wired. Wired is worth doing.

Your electricity

The other failure mode, and the more annoying one. An unstable supply takes down a machine that was mid-task, and the work in flight goes with it.

What makes it expensive is not the interruption. It is that the agents stop and nothing starts them again. There is no supervisor watching, so the machines come back and simply sit there. I find out because I go and look.

That is probably fixable — a launch agent, a health check, something that notices and restarts — and I have not fixed it yet. Until I do, every power event costs however long it takes me to notice.

I have not solved this yet, so treat what follows as reasoning rather than a report. Putting the Mac mini on a UPS is the obvious answer, and I think it is only half of one. If the power dips, the router and the modem go down with it, and an agent with a live machine and a dead connection is no better off than an agent with no machine at all.

So the two failure modes are really one purchase. A UPS has to carry the network gear as well as the machine, which changes the capacity you need and therefore what you buy.

I will write that up properly once I have done it rather than argued it.

Would I buy it again

Yes. The Mac mini has been a very good investment for this. It is quiet enough to live in a room you use, it does the job without complaint, and three of them cost less than the single large machine I might otherwise have talked myself into.

What I would change is the shopping list, not the machine. Less time agonising over memory tiers. More attention to storage headroom, a wired connection, and keeping the power on.

The short version

If your agents call APIs:

• Memory: 24GB is plenty. Do not pay up for a tier you will not use.
• Storage: 512GB internal if you can, and plan on external storage anyway.
• Count: more machines beats one bigger machine.
• Look at next: a wired connection to the router, and a UPS with enough capacity for the router as well as the machine. I have not bought one yet.

If you are running models locally, ignore all of this. You are solving a different problem, and memory really is your answer.

About the products below

I own the Mac minis. I do not own the UPS, the enclosure, or the switch. They are the specifications I reached by running this setup and watching what actually broke, and they are what I intend to buy.

I would rather say that than pretend otherwise. When I have used them, this section changes and so does what I recommend.