Apple's Claris Is Testing Harbor, a Private AI Server That Pools Memory Across Macs Over Thunderbolt 5. Here Is What It Would Replace in My Setup.
On November 10, a leaked post on X and then reports from MacRumors and 9to5Mac pointed to a page from Claris, "an Apple company," offering "Early access to a new private AI server." The product is called Harbor. It runs open-weight models on the Apple silicon Macs you own, and the line that got everyone's attention is this one: "Connect two or more Thunderbolt 5-capable Mac computers and Harbor pools their unified memory into a single inference resource." If you've been eyeing a Mac Studio as a local model box, this is the first sign Apple wants to own that use case. It is also a private preview with almost no specs, so let's separate the parts you can plan on from the parts you can't.
What is confirmed
From the Claris page itself: Harbor is labeled "Private Preview." It lets you "browse, download, and load open-weight models from within the app." It promises "zero cloud fees, no per-token billing," which describes running costs, not the price of the app. You can "set up an endpoint in minutes on a single Mac," and with multiple Thunderbolt 5 Macs it detects the devices and forms a cluster automatically.
From MacRumors: Harbor "is in early access and is not publicly available," and an App Store listing appeared and was then removed. The sign-up form asks for name, email, company, phone and country, which reads like a business pilot rather than a consumer beta. MacRumors also notes that RDMA over Thunderbolt 5 arrived in macOS Tahoe 26.2. That's the OS feature that makes memory pooling across Macs practical.
9to5Mac's report describes the same app and says a public launch "seems likely" in the coming weeks and months. That's their read, not a date.
What is not confirmed
Price: not stated anywhere. MacRumors says it isn't clear whether it will be free. Hardware minimums: only "Apple silicon," no chip generation, no RAM floor. Models: no names. Maximum cluster size: not stated. And the one I care most about as a builder: the API. "Endpoint" strongly suggests an HTTP server, but nothing says it speaks the OpenAI-compatible format that most of our tools expect. Until that's documented, I would not assume my existing code can point at it unchanged.
What it competes with
You can already do most of this today. LM Studio and Ollama run open-weight models on a single Mac and expose a local API. Exo, the open-source project MacRumors mentions, already clusters Apple silicon machines to run models too large for one of them. People have been chaining Mac Studios over Thunderbolt 5 to run very large models for most of this year.
So Harbor's pitch isn't a new capability. It's a first-party, supported, presumably polished version of what hobbyists wire up by hand. For a solo operator that matters in a specific way: when something breaks in a three-Mac Exo cluster at 11pm, you're on GitHub issues. When it breaks in an Apple-owned product, there's at least someone to email, and the next macOS update is less likely to break it.
The math to run before you buy hardware
The "no per-token billing" line is true and also incomplete. Local inference moves the cost from a monthly bill to a hardware purchase plus electricity plus your time. Here's how I'd frame it for a small product:
- Take your current monthly API spend for the tasks an open-weight model could actually handle. Not all your spend, only the part where a local model is good enough. For many of us that's classification, extraction, summarization and drafts, not the hard reasoning calls.
- Price the Mac you'd need to hold that model in memory, and divide by 24 or 36 months.
- Add a realistic number for your time keeping it running.
If the monthly API spend on that slice is $40, a multi-thousand-dollar Mac never pays back on cost alone. If it's $600, or if you have data you can't send to a third party, the case gets real quickly. The pooling feature only matters if the model you need doesn't fit in one machine, and for most solo workloads a single well-specced Mac covers it.
Where I could be wrong
There are two ways this goes better than I'm assuming. If Apple prices Harbor at zero, ships an OpenAI-compatible endpoint and certifies a short list of models, it becomes the obvious default for anyone with a recent Mac, and the time cost in my math shrinks to near nothing. And if Apple adds its own models to the catalog, the quality gap with hosted APIs could narrow for everyday tasks.
The way it goes worse: it's a Claris product aimed at businesses, with a company-name field on the sign-up, and it ends up as a per-seat enterprise subscription. That would be fine for Apple and irrelevant for most of us.
What I'd actually do
Sign up for the preview if you're curious. It costs nothing. Don't buy a Mac for Harbor until there's a price and an API spec. In the meantime, if local inference is on your roadmap, set it up now with Ollama or LM Studio on the Mac you already have, and route one low-stakes task through it for a month. You'll learn what model size you actually need and whether local is good enough for your workload. Those answers will still be true when Harbor ships, whatever it costs.
Author
Lukas
@lukcombinator