Apple is preparing Harbor app to run AI locally on a Mac cluster

What the leak shows

On November 10, 2026, leaker Aaron Perris published in a post on X evidence of an app not yet announced by Apple called Harbor. According to 9to5Mac, Harbor is in private preview under the Claris brand, Apple's software division historically linked to FileMaker. There is no official confirmation from Apple nor a public launch date.

The description found by Perris is direct about what the app proposes to do:

Harbor turns your Mac into a private AI inference server. Run open-weight models locally on Apple silicon—your data never leaves your hardware, and nothing reaches the cloud. Connect multiple Mac computers over Thunderbolt 5 to pool unified memory, enabling large models and higher concurrency that a single computer can handle alone.

Aaron Perris, leaker, via X, quoted by 9to5Mac

Setup with no terminal, models inside the app

The leaked material describes three fronts of functionality. The first is automatic detection: Harbor would recognize other Macs connected via Thunderbolt 5 and form a cluster on its own, without requiring manual configuration, command line, or infrastructure knowledge. The second is model management: the app would allow browsing, downloading, and loading open-weight models directly from the interface, removing from the user the step of handling weights, formats, and dependencies by hand.

The third front is clustering itself: by connecting two or more Macs with Thunderbolt 5, Harbor would pool the machines' unified memory into a single inference resource, allowing models that exceed the capacity of an isolated computer to run, or serving more simultaneous requests distributed across the machines.

Why this matters for those already running local models

Anyone who has already tried running a large language model locally knows that the friction is rarely in the model itself: it's in the surrounding setup. Today, running something like a 70B or larger model on Apple silicon hardware goes through tools like Ollama, LM Studio, or Apple's own MLX framework, and scaling beyond one machine usually means manually configuring distributed modes (like llama.cpp's RPC) or turning to community projects like exo, which already proposes joining multiple Macs over a network to serve larger models.

What Harbor promises, according to the leaked description, is to package exactly that annoying part: node discovery, cluster formation, and memory pooling, without the dev needing to write network configuration or touch a terminal. For those evaluating inference infrastructure on a small team, this changes the cost calculation: instead of spinning up a cloud GPU for a model that doesn't fit in a Mac Studio's memory, the idea would be to plug in a second Mac via Thunderbolt 5 and gain the combined unified memory.

In short: Harbor's pitch is to compete directly with the operational complexity that today separates