Apple’s Claris subsidiary is testing an early-access app, Harbor, that would turn an Apple silicon Mac into a private AI server. You’d download open-weight models in the app, run them on your own hardware, and optionally cluster Thunderbolt 5 Macs to pool their memory. There’s no public release date, and a brief App Store listing has already been taken down.
What Harbor is meant to do
According to Claris’s own early-access page, Harbor “turns the Apple silicon computers you own into a private AI server.” It says you can run open-weight models with “zero cloud fees, no per-token billing, and the assurance that your data never leaves your environment.” You can start on a single Mac, then connect more machines over Thunderbolt 5 to pool unified memory.
Marketing copy shared this morning (11 November, Melbourne time) by Aaron Perris on X says you’d browse, download and load models in the app, with no command line. Two or more Thunderbolt 5 Macs would be detected automatically and treated as one inference resource, so you can run models that don’t fit on a single machine, or handle more requests at once. The materials only mention Macs. There’s been no report of an iPhone or iPad version.
Spotted, then pulled from the App Store
Perris, a MacRumors contributor, posted screenshots of the pitch and an App Store listing. He later confirmed the project was previously called Fabric, after a download path still referred to com.claris.fabric. By mid-morning he said the listing had been taken down. The Claris sign-up page is still live.
Neither Claris nor Apple has issued a formal launch announcement, or given a release date or price. The sign-up form requires a company name, which may point to a business focus, though that hasn’t been confirmed.
Not Siri, and not the first way to do this
Harbor is a different idea from Apple Intelligence and Siri AI, which use Apple’s own models, some on the device and some through Private Cloud Compute. This is about running third-party, open-weight models on hardware you control. For example, I I use the Parakeet V3 model in MacWhisper for transcriptions and Ollama as an on-machine alternative to tools like ChatGPT.
When Apple announced the M5 Max and M5 Ultra Mac Studio in August, it talked up clustering over Thunderbolt 5 and RDMA, along with MLX, its open-source Apple silicon framework, and Core AI, a new framework for running models locally. A WWDC26 session on local agentic AI with MLX noted that Ollama, LM Studio and vLLM already build on that stack.
Mac users can already do much of this. MacRumors notes that LM Studio covers the single-Mac case and the open-source tool Exo can cluster Macs, and that Harbor looks like an attempt to put a simple interface over both. None of the coverage says Harbor itself is built on MLX.
The catch is still the hardware
Clustering, as described, only works with Thunderbolt 5-capable Macs. That’s the current Mac Studio, and some other recent models — the base Mac mini with M6 uses Thunderbolt 4, so it wouldn’t join a Harbor cluster. Memory is the other limit. A Mac Studio with M5 Ultra can take up to 512GB of unified memory. In Australia it starts at $4,299 including GST for the M5 Max and $9,499 for the M5 Ultra, before you buy a second machine to cluster.
If prompts and data really stay on your own Macs, that’s the appeal for anyone who’d rather not send work to a cloud model. For now, Harbor is early access only, with no date for a public release.
Anthony is the founder of Australian Apple News. He is a long-time Apple user and former editor of Australian Macworld. He has contributed to many technology magazines and newspapers as well as appearing regularly on radio and occasionally on TV.