Apple is reportedly preparing Harbor, an app to turn your Macs into a private AI server
The gold rush < pickaxes
Stacked Mac minis on a desk, connected by cables, to run a home-grown AI model: the sight has become familiar in 2026, and Apple seems to have decided to turn it into a product. Aaron Perris, an analyst for MacRumors with a track record of leaks, revealed on X on November 10 an application named Harbor, described as a “private AI inference server.”
The app is in private preview under the Claris brand, Apple’s subsidiary best known for its FileMaker database software (does that still exist?). According to MacRumors, an App Store listing briefly existed before being removed.
Harbor, the port for welcoming other people’s boats
Harbor is used to download and run open-weight models (those with public weights, such as Mistral’s and many others’) directly on a Mac equipped with an Apple Silicon chip. Everything happens in the app: you browse a catalog, download the model, and load it. Requests stay on the machine; nothing is sent to the cloud.
Harbor capable of managing multiple Macs // Source: MacRumors
The real innovation lies in cluster mode.Connect two Macs compatible with Thunderbolt 5 or later: Harbor automatically detects them and pools their unified memory into a single computing resource. Enough to load models too large for a single machine, or handle more simultaneous requests. The description promises setup in just a few minutes, with no command line or manual configuration.
In the published screenshots, a MacBook Pro M5 Max (48 GB) is connected to two Mac mini M4 Pro models (16 and 24 GB), OpenAI’s gpt-oss-20b model is loaded, and there is a button for copying the address of a local API. In other words, other software could query this small home server just as it queries ChatGPT.
Why Apple is interested
The Mac has become, somewhat by accident, a machine of choice for local AI. Its unified memory, shared between the CPU and GPU, makes it possible to load large models without buying a professional graphics card (if you’re wondering whether your computer is capable, we explain how to check). Apple also added RDMA over Thunderbolt 5 in macOS Tahoe 26.2, which speeds up exchanges between Macs in a cluster, and unveiled new Mac mini and Mac Studio models designed for local AI in late August.
Harbor isn’t inventing the wheel, though: the open-source Exo tool already makes it possible to connect multiple Macs, and LM Studio makes things simple on a single machine. According to MacRumors, Harbor looks like a mix of the two, with Apple’s polished finish as a bonus (and without spending your evening in Terminal).
+ fast, + practical, + exclusive
No ads, advanced reading features, articles summarized by A.I., exclusive content, and more.
Discover the many benefits of Numerama+.
You have read 0 articles on Numerama this month
Not everyone can afford to pay for information.
That's why we keep our journalism open to everyone.
But if you can,
here are three good reasons to support our work:
- 1 Numerama+ helps provide a free experience to all Numerama readers.
- 2You will enjoy an ad-free reading experience, numerous advanced reading features and exclusive content.
- 3 Support Numerama in its mission: understand the present to anticipate the future.
If you believe in a free web and quality information accessible to as many people as possible, join Numerama+.
All the tech news at a glance
Add Numerama to your home screen and stay connected to the future!
To not miss any of the latest news, follow Numerama on Google!
To read
Anticipate the future with Numerama’s Always+ newsletter!
1,147 pages of “daunting” complexity: physicists refuse to publish OpenAI’s proof too soon despite the mathematicians’ call
“No scientific reason to delay” the proof: mathematicians are up in arms over OpenAI’s stance on a crucial physics result
OpenAI claims to have solved one of the greatest mysteries in physics, but keeps the evidence secret
Containing AI is becoming a headache: after its 11-day pause, OpenAI explains why cutting off internet access is not enough
AI giants are already preparing for the aftermath of the disaster and particularly dread two scenarios
