Claris, the company behind Apple's database software FileMaker, is testing an app called Harbor. The app turns Macs running Apple silicon into a dedicated AI server, running open-weight models locally without ever moving data off the device.

What happened?

Harbor can run on a single Mac, but its boldest feature is linking multiple Macs together over Thunderbolt 5 and merging their unified memory into a single shared pool. Apple appears to have sped up this clustering by adding RDMA support over Thunderbolt 5 in macOS Tahoe 26.2. The app automatically detects devices, manages model downloading and loading from a single screen, and requires no command line. It isn't publicly available yet; an App Store listing was spotted but then pulled, and the sign-up page asks for a company name.

Why it matters?

Apple's AI strategy has followed the same line for two years now: run what you can on-device, and even when the cloud is needed, keep the data on infrastructure it controls itself. Having Apple Intelligence handle part of its workload on-device and offload the heavy lifting to its own servers under the name Private Cloud Compute was the first major expression of that approach. Harbor points the same philosophy at enterprises rather than consumers this time: the ability to build a closed, in-house AI server without depending on OpenAI's or Google's cloud APIs.

The hardware side that makes this possible is no accident either. Apple silicon's unified memory architecture lets the CPU, GPU, and Neural Engine share the same memory pool; this memory, which is critical for large language models, can exceed a hundred gigabytes on top-tier Mac Studio configurations. Pooling the memory of multiple Macs over Thunderbolt 5 brings — on a smaller scale — the kind of work done with GPU clusters in data centers to desktop hardware. Nvidia's DGX line in enterprise AI infrastructure solves exactly this need with GPU clusters; Apple's answer is to lean on the memory advantage of its own chip architecture.

Who does this affect, and how?

Harbor's real target is probably not individual developers but organizations sensitive about data privacy. The fact that the sign-up page asks for a company name, and that Claris already sells FileMaker to business customers, supports that reading. But the market gap isn't huge: LM Studio has already made it simple to run open models on a single Mac, and Exo has offered open-source Mac clustering for a while now. What sets Harbor apart is combining both of those functions into a single, branded, and presumably supported product.

ToolSingle-Mac supportMulti-Mac clusteringEnterprise focus
LM StudioYesNoLow
ExoYesYes (open source)Low
HarborYesYes (Thunderbolt 5)High (claimed)

Pricing, and whether it will be free, remains unclear — which fits with Claris's history of enterprise licensing. It would be no surprise at all if it arrives as a paid, managed piece of software.

What's next?

The fact that the app is still in an early-access stage and that its App Store listing was pulled shows Apple hasn't yet matured the product. Three things are worth watching: when an official announcement will come, whether the pricing model will also cover individual users, and how large the supported model pool will be. Crowning an area where Apple's hardware is already strong — high-memory Macs — with software is a concrete sign of the company's ambition to offer an alternative to cloud providers in AI infrastructure. Even so, this isn't a serious short-term threat to Nvidia's position in the enterprise market; it's more likely to open up a niche for mid-sized organizations that prioritize privacy.