Apple's Claris subsidiary is testing Harbor, an app that turns Apple silicon Macs into a private AI server, including clusters of several Macs wired together.
Harbor works on a single Mac, but the feature that stands out is linking multiple machines over Thunderbolt 5 to pool their unified memory, the shared memory Apple's chips use for both processor and graphics work. More pooled memory means larger models fit. The app's description also promises higher concurrency, meaning more simultaneous requests, than one computer could handle alone. Apple added RDMA over Thunderbolt 5 in macOS Tahoe 26.2, which makes clustering faster. RDMA lets machines exchange data memory-to-memory with less overhead.
Harbor detects connected devices on its own, then lets users browse, download and load open-weight models, meaning models whose trained parameters are publicly downloadable. The description says setup needs no command line and no manual configuration, and that data never leaves the user's hardware. Those are the company's claims, and the source offers no benchmarks to test them against.
Running models locally on a Mac is not new, and model management is not what separates Harbor. LM Studio, for one, already lets users browse and download models from inside its app. The real difference is multi-Mac clustering: Thunderbolt 5 links, operating-system support for RDMA, and a pitch of plug in and start.
Harbor is in early access and not publicly available. An App Store listing appeared, then was removed. Claris is best known for FileMaker, a database platform aimed at businesses, and Harbor's sign-up page asks for a company name. How Apple will market the app, and whether it will be free, remains unclear.