Apple’s Claris previews Harbor, an app that runs local AI across Macs. What works now

What was found on November 10

On November 10, 2026, the leaker Aaron Perris posted on X that “Apple is working on a new app called Harbor,” with the app’s own description, ending: “Harbor is currently in a private preview under the Claris brand.” 9to5Mac reported it within the hour.

MacRumors then linked the page behind it. Claris, the Apple subsidiary known for its FileMaker database, has a sign-up page for Claris Harbor, labeled “Private Preview.” MacRumors adds that “There was an App Store listing, but it has now been removed,” and Cult of Mac writes that “Apple took down the App Store listing” after Perris spotted it. AppleInsider says the app was “Formerly known as Fabric,” and the header image on Claris’s page still carries the text “Claris Fabric” in its code.

So this is a leak with a company page behind it. Here is who says what:

Claim	Who says so
Harbor exists, as a private preview from Claris	Claris’s own page
It serves open-weight models from one Mac, or several over Thunderbolt 5	Claris’s own page
It finds Thunderbolt 5 Macs by itself, and downloads models inside the app	The app’s description, as Perris posted it
Its App Store listing was removed	MacRumors and Cult of Mac
It used to be called Fabric	AppleInsider, and the image text on Claris’s page
It will launch “in the weeks and months ahead”	9to5Mac’s expectation, not Claris’s

What Claris’s page says

The page is titled “Claris Harbor | Claris, an Apple company,” and offers “Early access to a new private AI server from Claris, an Apple company.” Its pitch:

• “Claris Harbor turns the Apple silicon computers you own into a private AI server. Run open-weight models with zero cloud fees, no per-token billing, and the assurance that your data never leaves your environment.”
• “Set up an endpoint in minutes on a single Mac. When you want to run larger models, connect multiple machines over Thunderbolt 5 to pool unified memory on hardware you control.”
• “Data and prompts stay strictly within your physical Mac hardware.”

The description Perris posted, which 9to5Mac and MacRumors quote, goes further on three points:

• Setup. “Harbor automatically detects your devices over Thunderbolt 5 to form a cluster. Plug in, start, and you’re running—no manual configuration, no command line, no infrastructure expertise required.”
• Models. “Browse, download, and load open-weight models from within the app. Harbor handles model management so you can focus on the work, not the setup.”
• Clusters. “Connect two or more Thunderbolt 5-capable Mac computers and Harbor pools their unified memory into a single inference resource. Run models that exceed any single computer’s capacity, or serve more concurrent requests across your cluster.”

The form on Claris’s page reads “Complete the form to stay informed” and “Fill out your details below to receive updates.” It asks for a first and last name, email, company, phone, country and comments. The heading says early access, but the form itself promises only updates. The page doesn’t say who gets the app, or when.

What Claris and Apple haven’t said

• A price. MacRumors: “it isn’t clear how Apple will market the Harbor app or whether it will be free. The sign-up page for Harbor asks for a company name, which suggests more of a business focus.”
• A date. The page has none. 9to5Mac’s “weeks and months ahead” is a guess, and AppleInsider’s is the opposite: “End users shouldn’t expect to see the app anytime soon.”
• Requirements. The page names no macOS version, no memory size and no maximum number of Macs.
• Which models it offers, and what its endpoint speaks. The page says “endpoint” and stops there, so whether existing tools can connect to it is unknown.
• Anything from Apple itself. As of November 11, we found no mention of Harbor in Apple’s Newsroom or on Claris’s homepage and blog. None of the four reports we read quotes Apple or Claris.

What does the same job today

Most of what Harbor promises can be done now, with more work. MacRumors names two tools: “Exo is an open-source tool for clustering Macs, and LM Studio is available for single-Mac use, but Apple’s Harbor app appears to combine LM Studio’s simple interface with clustering.” Apple’s own MLX is the third.

Way	Several Macs	Setup	Status on November 11, 2026
Harbor, per Claris and the leak	Over Thunderbolt 5, found automatically	In the app, with “no command line”	Private preview; no price or date
MLX, Apple’s open-source tools	mlx.launch with a hostfile, over Thunderbolt or Ethernet	Command line, with ssh between the Macs	Available now
exo	Found automatically; RDMA over Thunderbolt 5 on macOS 26.2 or later	A menu bar app with a dashboard	Available now, open source under Apache 2.0
LM Studio	“Single-Mac use,” per MacRumors; it serves “locally and on the network”	An app that downloads models and serves them	Available now

MLX. Apple’s WWDC26 session Run local agentic AI on the Mac using MLX calls MLX “our open-source array framework purpose-built for Apple silicon” and says “MLX’s distributed support lets you spread a model across multiple Macs connected over Thunderbolt or Ethernet.” Its companion session joins four M3 Ultra Macs and reports “nearly 3x the token generation rate of a single machine” on a 27-billion-parameter Qwen model, plus “a demo running the 1-trillion-parameter Kimi 2.6 model across four Macs.”

exo. Its README says “exo connects all your devices into an AI cluster” and “Devices running exo automatically discover each other - no manual configuration.” The Mac app “needs a Mac with Apple silicon and macOS Sequoia 15.1 or later,” and exo claims “up to 1.8x speedup on 2 devices and 3.2x speedup on 4 devices.” Its API is “Compatible with OpenAI Chat Completions API, Claude Messages API, OpenAI Responses API, and Ollama API.”

LM Studio. Its docs say you can search for and download models inside the app and “Serve local models on OpenAI-like endpoints, locally and on the network,” and that “On Apple Silicon Macs, LM Studio also supports running LLMs using Apple’s MLX.” Apple’s session counts it among the tools built on MLX: “Ollama, LM Studio, and vLLM are just a few of the most popular ones.”

Which Macs can join a Thunderbolt 5 cluster

Apple’s ports page lists the Macs with Thunderbolt 5:

• Mac Studio introduced in 2025 or later
• 14-inch MacBook Pro with Pro or Max chip introduced in 2024 or later
• 16-inch MacBook Pro introduced in 2024 or later
• Mac mini with Pro chip introduced in 2024 or later

The fast link between them is newer. The macOS 26.2 release notes add RDMA over Thunderbolt, which “Enables low-latency communication between Thunderbolt 5 hosts for use cases including distributed AI inference using MLX.” MLX’s docs say it brings “communication latency an order of magnitude lower” than MLX’s ring backend, which uses TCP sockets.

It takes some setting up today, which is the work Harbor says it removes:

• Turning it on. MLX’s docs say “enabling RDMA over thunderbolt is slightly more involved and cannot be done remotely even with sudo”: you start the Mac in Recovery, open Terminal and run rdma_ctl enable. Apple’s WWDC26 session instead describes “enabling RDMA in System Settings.”
• Cabling. MLX’s docs say “there needs to be a thunderbolt cable connecting all pairs of Macs directly.” exo adds that “The cables must support TB5” and “On a Mac Studio, you cannot use the Thunderbolt 5 port next to the Ethernet port.”
• Matching software. exo warns that RDMA ports “may be unable to discover each other on different versions of MacOS,” and that Macs on different versions of exo don’t join each other.

How to serve a model from your Mac now

Apple’s session gives three steps: “install MLX-LM,” “start the server,” and “point your agent at the local server.”

pip install mlx-lm
mlx_lm.server --model mlx-community/Qwen-3.5-4B-8bit

The server answers at http://127.0.0.1:8080/v1 and is, in Apple’s words, “an OpenAI-compatible HTTP server.” mlx-lm’s own page warns that it “is not recommended for production as it only implements basic security checks,” so keep it on your own computer. For several Macs, Apple’s sample wraps the same server in mlx.launch:

mlx.launch --hostfile hosts.json \
  --backend jaccl \
  /remote/path/to/mlx_lm.server \
  --model mlx-community/Qwen-3.5-122B-A3B-8bit

The agent on top can be, Apple says, “Xcode, OpenCode, Pi agent, a custom script, or anything else,” and its session includes an OpenCode configuration pointed at that address. OpenCode’s provider docs show how to add any OpenAI-compatible endpoint by its base URL, and say “LM Studio and vLLM also have built-in local discovery.”

Enjoy runs OpenCode beside Claude Code, Codex and the other labs’ agents, and its model picker lists the models OpenCode has set up, local ones included. So a model served from your Mac, or from a cluster of them, can take over a conversation when a plan runs low. Enjoy doesn’t serve the model: MLX, LM Studio or exo does, and setting that up happens in OpenCode. Whether Harbor’s endpoint will work the same way is one of the things Claris hasn’t said. Enjoy is a monthly subscription, with the first seven days free (pricing). Our guide to coding agents on local models covers memory, context and the other agents that can do this.

What to do now

1. Sign up if you want the news. Claris’s form is the only official page we found, and it promises updates.
2. Don’t buy Macs for Harbor yet. Claris has published no requirements, price or date.
3. Don’t install anything calling itself Harbor. MacRumors reports that it “is not publicly available,” so treat any download offered under that name as someone else’s.
4. On one Mac, start now. LM Studio or mlx_lm.server gives you a local endpoint today.
5. On several Macs, try exo or MLX. Check the Thunderbolt 5 list above first, and update every Mac to the same macOS.

We’ll update this memo when Claris or Apple says more.