Apple's New App Harbor: Connect a Few Macs Together to Run Local Large Models

Apple is developing a new app called Harbor. The news comes from screenshots and a product description Aaron (@aaronp613) posted on X. Its goal is straightforward: turn a Mac into a private AI inference server that runs open-weight models locally, with data not leaving the hardware and not going to the cloud.

Harbor's product description lays out several key points. Models run locally on Apple silicon; multiple Macs can be connected via Thunderbolt 5 to pool their unified memory into a single inference resource; this makes it possible to run models that don't fit in a single machine's memory, or to handle higher concurrency. Setup is compressed into a few minutes, and the app automatically discovers Thunderbolt 5 devices and forms a cluster, with no manual configuration, command line, or infrastructure experience required.

For readers unfamiliar with hardware, unified memory can simply be understood as the CPU and GPU sharing the same memory. Pooling unified memory across multiple Macs is like combining the memory of several machines into one larger pool. Thunderbolt 5's bandwidth is high enough to make this kind of pooling feasible on desktop-class devices. Harbor also includes built-in model management, letting users browse, download, and load open-weight models within the app.

What's worth noting here isn't 'running models locally' itself. llama.cpp, Ollama, and LM Studio can all do that. What Harbor is really trying to solve is the memory wall and cluster configuration: large models eat memory, and even the most expensive single Mac has a ceiling; multi-machine inference usually requires fiddling with networking, model sharding, and scheduling. If Apple can turn 'plug in Thunderbolt 5, launch, and run' into a product, the barrier to entry will be completely different.

Harbor is currently in private preview under the Claris brand. The name Claris has appeared in Apple history before, and some commenters immediately posted memes about it. A more detailed clue: a user saw com.claris.fabric in the download path in the second screenshot, and Aaron replied that it used to be called Fabric before being renamed Harbor. The product isn't public yet, but traces of the rename are already left in the path.

There's also a demo video in the replies:

The comments section hasn't been idle either. Some mention that Exo has been doing multi-machine inference for a while, and others ask whether Exolabs is involved in Harbor, but there's no confirmation. Others noticed that M2 Pro appears in the promotional material, while M2 Pro doesn't support Thunderbolt 5. Harbor's clustering feature explicitly depends on Thunderbolt 5, and this detail isn't explained in the original post. Is the ad material using old images as filler, or does Harbor have another plan for non-Thunderbolt 5 devices? For now, we can only wait for more information.

My take: Harbor's significance lies in turning 'cluster inference' from a geeky manual chore into a consumer-grade app. Apple isn't piling up GPUs in the cloud; it's pulling compute back to the Macs on your desk. The direction isn't new, but with Apple doing it, the productization and default experience will be different. Regular users shouldn't rush to calculate electricity costs just yet: it requires multiple Macs that support Thunderbolt 5, the cost isn't low, and it's still in private preview. Developers and teams with small Mac clusters should keep a close eye on it. If it really can form a cluster in minutes, the way local inference is done will change.

Original tweet link: https://x.com/aaronp613/status/2120229190477992162