Apple reportedly building Harbor, a Mac app for clustered local AI inference

A leak points to a new Apple tool for local AI

Apple is working on a Mac application named Harbor that would run open-weight AI models directly on Apple silicon, according to 9to5Mac. The report draws on findings from leaker Aaron Perris, who described the software as a way to turn a Mac into a private inference server, one where prompts and data stay on the machine and nothing is handed off to the cloud.

What Harbor reportedly does

Perris's description, relayed by 9to5Mac, lays out three core capabilities:

• Simple setup: Harbor would automatically detect other Thunderbolt 5-capable Macs and link them into a cluster, with no manual configuration, command-line work or infrastructure expertise required.
• Built-in model management: users could browse, download and load open-weight models from inside the app, instead of wrangling model files with external tools.
• Clustering for scale: with two or more connected Macs, Harbor would merge their unified memory into a single inference resource, allowing a cluster to run models too large for any one machine to hold, or to serve more requests simultaneously.

The Thunderbolt 5 detail is the most striking part of the leak. If it works as described, a pair of Mac minis or a rack of Mac Studios could behave as one larger inference machine, leaning on the sizable unified memory pools that Apple silicon offers relative to typical discrete GPUs.

Where it comes from

Harbor is currently in a private preview and is being developed under Apple's Claris brand rather than the main Apple name, 9to5Mac reports. The outlet expects a public release in the weeks or months ahead, though neither Apple nor Claris has announced anything, and the details come from a single leaker's account.

The broader Mac picture

9to5Mac frames the app against a year in which the Mac has grown more popular, partly through lower-priced machines and partly through demand for running AI models locally on hardware such as the Mac Studio and Mac mini. Harbor would formalize that use case with a first-party tool rather than leaving users to community projects and command-line utilities.

Why it matters

Local inference has largely been the territory of open-source tooling, with Apple supplying underlying frameworks and APIs but little consumer-facing software for running arbitrary open-weight models. A native app would signal that Apple treats local AI serving as a mainstream use for its hardware, not a niche hobbyist pursuit.

The privacy pitch also appears central to the design as described. If inference never leaves the hardware, the app would suit organizations that want AI capabilities without sending sensitive material to a cloud provider, a positioning that lines up neatly with Apple's wider privacy messaging.

The memory-pooling idea may be the most consequential piece. Multi-machine clustering of this kind is genuinely hard to replicate elsewhere, and if Apple ships it, the software would tie directly into the architectural strengths of its own silicon, giving buyers a reason to stack Macs rather than reach for a GPU server.

None of this is confirmed, and features, branding or timing could change before any launch. But the leak suggests Apple is preparing to compete head-on with the ecosystem of local-model tools that has grown up around its own hardware.