Apple’s Claris Is Testing Harbor, an App That Pools Macs for Local AI

Running a big AI model on your own computer usually comes down to one number: how much memory you have. Apple looks ready to make that number bigger, by letting several Macs act as one. Harbor is a new app from Claris, a software company Apple owns. It turns Apple silicon Macs into a private AI server, and it can pool the memory of two or more Macs joined over Thunderbolt 5. It isn’t out yet. Claris has a sign-up page for updates, and the full feature list came from a leaked listing.

What is Harbor?

It’s a private AI server in an app. Your own. You run open-weight models, the ones anyone can download, on your own Mac, so your prompts and data never leave your hardware. Claris pitches it with no cloud fees and no per-token billing. That’s the part of local AI people like, minus the setup that usually comes with it.

The leaked listing, quoted by 9to5Mac and MacRumors, fills in the rest. Harbor lets you browse, download and load models from inside the app. It finds your other Thunderbolt 5 Macs on its own and builds the cluster for you. “No manual configuration, no command line, no infrastructure expertise required,” it says. If that holds up, it’s a big step. Getting two machines to share one model is usually the hard part.

How does pooling Macs together work?

It works because of how Macs handle memory. On Apple silicon, the CPU and graphics share one pool, called unified memory, so a Mac with lots of it can load a model that wouldn’t fit on most graphics cards. Harbor’s idea is to treat the memory of two or more Macs as one big pool. That lets you run a model too large for any one of them, or serve more requests at once.

Apple already laid the groundwork. macOS Tahoe 26.2 made Thunderbolt 5 connections between Macs much faster for this kind of job, and spreading AI work across machines was one of the uses Apple named. What’s been missing is an app that makes it easy.

There’s a catch. The port. Clustering needs Thunderbolt 5, which means a recent Mac with a Pro, Max or Ultra chip. If yours is a MacBook Air, an iMac or a base Mac mini, Harbor should still run on that one Mac, going by Claris’s page, but it can’t join a cluster. Check your ports before you plan on pooling anything.

What’s confirmed, and what isn’t?

Harbor is real, which is more than most leaks can say. Claris has a public page for it, marked private preview, and that page backs up the basics. The finer details, like the no-command-line setup and the model downloads, come from a listing MacRumors says was on the App Store before it was pulled. It reads like a finished product, which is a good sign, but Claris hasn’t put those details on its own page yet.

Then there’s the audience. Claris makes FileMaker, a database tool aimed at businesses, and the sign-up form asks for your company. That hints at a business product first. It could still be free, or it could be a subscription, and nobody knows yet.

Why does Harbor matter if you run AI at home?

Memory. It’s the wall everyone hits. It’s the same trade I looked at with AMD’s Ryzen AI Max+ PRO 495, where the jump to 192GB is mostly about fitting bigger models in one box. Harbor’s answer is to add a second Mac instead of buying a bigger one. Nvidia’s DGX Spark and boxes built on Axera’s AX8860 chip hold 128GB each, and a single Mac Studio with M5 Ultra already goes to 512GB. Pooling a few of those is where it gets interesting for really big models.

You can already do this today. Sort of. Exo is a free, open-source tool that clusters Macs, and LM Studio runs models on a single Mac. Exo already finds other Macs on its own. What Harbor seems to promise is that plus model management in one app, with Apple’s polish. If Claris prices it for home users, that’s the version I’d want to see.