Claris Harbor: Apple's Mac AI Server App, Explained

Claris Harbor is a coming Mac app that turns Apple silicon Macs into a private AI server. Claris, the Apple company that makes FileMaker, confirmed it on its own early-access page. Harbor runs open-weight models, which are AI models anyone can download and run. It can use one Mac, or pool the unified memory of several Macs linked by Thunderbolt 5 cables. Unified memory is the one pool of memory that an Apple silicon chip’s CPU and GPU share. You can’t download Harbor yet. The page calls it a “Private Preview,” and an App Store listing that showed up on November 10, 2026 was pulled the same day.

Harbor is aimed at people who want a local AI server without the command line. If you already split models across Macs with exo (a free cluster tool) or MLX (Apple’s machine learning framework), nothing changes for you today.

The news posts skipped one detail. The cluster screenshot shows RDMA as “Not Enabled,” next to a link labeled “Enable in Privacy & Security…”. RDMA (remote direct memory access) lets one Mac read another Mac’s memory with very little delay. Apple’s own docs say you turn it on from macOS Recovery. How Harbor’s link works is unknown.

For what fits on one Mac today, and how fast, use My Rig. It covers single Macs only.

Confirmed vs reported

Confirmed means Claris or Apple published it. Reported means it comes from the pulled App Store listing or from other people’s posts. Times are EST.

Aaron Perris, who also contributes to MacRumors, spotted the listing first. He said his post copied Apple’s listing text word for word, except his first and last sentences. MacRumors, 9to5Mac, AppleInsider and The Mac Observer covered it next. Some posts call it Apple Harbor. The official name is Claris Harbor.

What Harbor does

Harbor’s main idea is to add up unified memory across Macs. The Claris page puts it this way: “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.”

An endpoint is a network address that other apps send requests to. The API is the set of requests it accepts. So the Harbor app for Mac does three things, per Claris and the pulled listing:

1. Runs a model on one Mac and serves it to other apps on your network.
2. Joins two or more Thunderbolt 5 Macs into a cluster. The listing says Harbor “automatically detects your devices over Thunderbolt 5 to form a cluster.” It says a cluster can run models too big for one Mac, or serve more requests at once.
3. Lets you browse, download and load open-weight models inside the app.

What the screenshots show

These come from the App Store listing images and from Aaron’s own screenshot. They are reported, and mock-up values in promo images are not specs.

• A cluster view links a MacBook Pro (M5 Max, 48 GB) to two Mac minis (M4 Pro, 16 GB and 24 GB) with cables marked “TB5.”
• A side panel shows Server Status “Running,” Mode “Standalone,” an API Endpoint with a local network address and a “Copy API” button.
• Under Hardware, the panel shows Thunderbolt “3 ports” and RDMA “Not Enabled,” with the “Enable in Privacy & Security…” link.
• The loaded model is “gpt-oss-20b-MXFP4-Q4.” A download dialog says “Downloading from HuggingFace…” from the mlx-community/gpt-oss-20b-MXFP4-Q4 repo.
• Models save to a folder under ~/Library/Application Support/com.claris.fabric/. That path still uses the old Fabric name.
• There is a “Model Sandbox Chat” pane for testing a model.
• Aaron’s own screenshot shows the app open on his M2 Pro MacBook Pro with 16 GB. It says “Connect a Thunderbolt cable to another Mac” and has a “Dev Tools” item. He said he had not tested it yet.

Here is our reading, which Claris has not confirmed. The sample model is an MLX build from the mlx-community group on Hugging Face. That suggests Harbor runs models with MLX. Claris’s current FileMaker product, Claris AI Model Server, already uses MLX models on Macs. The API endpoint suggests other apps can call Harbor. Whether it speaks the same format as OpenAI’s API is unknown.

How to get early access

The Claris page has a form titled “Stay informed on Claris Harbor.” It asks for your first and last name, email, company, phone, country and comments. There is no download link. Claris has not said who gets in or when. MacRumors points out that asking for a company name suggests a business focus.

There is no Harbor macOS app to download from the App Store today. Claris offers no download, so treat any “Harbor Mac download” link elsewhere with care.

Which Macs can join a Harbor cluster

Claris says Harbor clusters Macs over Thunderbolt 5. It has not published a list of supported Macs. The table below lists the current Macs that have Thunderbolt 5 ports, from Apple’s spec pages, checked November 11, 2026. Having Thunderbolt 5 does not confirm Harbor support.

Some older Macs also have Thunderbolt 5. The exo README names the M4 Pro Mac mini, M4 Max Mac Studio, M4 Max MacBook Pro and M3 Ultra Mac Studio. The exo FAQ adds every current Mac Studio, the Mac mini with M4 Pro or M5 Pro, and every current MacBook Pro with a Pro or Max chip.

These current Macs have Thunderbolt 4 only, per Apple’s spec pages: the M6 Mac mini, the M6 MacBook Pro, the M6 iMac and the M5 MacBook Air. Older Macs with base M1, M2 or M3 chips, and the M2 Pro, lack it too. Going by Claris’s wording, none of these can join a Thunderbolt 5 cluster. Whether Harbor runs on them as a single Mac is unknown.

So a Mac mini cluster over Thunderbolt 5 means M4 Pro or M5 Pro minis. A Mac Studio cluster over Thunderbolt 5 can use any current Mac Studio.

What pooled memory means for model size

Adding up memory across Macs gives a bigger total on paper. You can’t use all of it as one block. Each Mac keeps some memory for macOS, and the model is split across the machines. Each Mac also needs room for the KV cache, which is the model’s saved context from your conversation, and for the runtime that runs the model.

So a model file smaller than your memory is only the first check. It does not tell you the model will run. Q4 means the model’s numbers are compressed to about 4 bits each. The table compares Q4 file sizes from our catalog with the most memory one Mac can have today:

• 64 GB: the top Mac mini M5 Pro or MacBook Pro M5 Pro.
• 128 GB: the top M5 Max.
• 512 GB: the top M5 Ultra.

Read the table this way. A “Yes” means the file alone is bigger than that Mac’s memory, so one such Mac can’t hold it. A “No” means only that the file is smaller. Our sizes are for GGUF files, the format llama.cpp uses. MLX builds of the same model, which Harbor appears to use, can be a different size. The table makes no claim that a cluster runs these models, or how fast.

For one Mac, My Rig checks fit with room for context and estimates speed. Try the M5 Max with 128 GB or the M5 Ultra with 512 GB. Our system pages cover the Mac Studio M5 Ultra (96 GB), Mac Studio M5 Max (64 GB) and Mac Studio M5 Max (36 GB). We don’t estimate clusters.

The model in Harbor’s screenshot, gpt-oss-20b, is small. Our catalog lists its Q4_K_M file at 11.67 GB, so one Mac is enough for it. Its page lists the Macs it fits.

Harbor has no published speed numbers. The Mac cluster results we know of used exo and llama.cpp on M3 Ultra Mac Studios. None used Harbor. Our Mac Studio cluster guide covers them.

What a Mac cluster needs today

Apple documents RDMA over Thunderbolt in technote TN3205, first published March 19, 2026 and updated April 13, 2026. This is what an RDMA Thunderbolt 5 Mac setup needs today, with any app:

• macOS 26.2 or later on an Apple silicon Mac with Thunderbolt 5. Apple released macOS 26.2 on December 12, 2025.
• RDMA turned on in macOS Recovery. Open Utilities, then Terminal, run rdma_ctl enable and restart. Check it with ibv_devices.
• Every Mac cabled to every other Mac for the lowest delay. This layout is called a full mesh. The links are point to point.
• Thunderbolt Bridge turned off when the cables form loops. Thunderbolt Bridge is the macOS setting that networks Macs over Thunderbolt.
• Idle sleep turned off, because Thunderbolt doesn’t support Wake on LAN. Apple also suggests automatic login and “Start up automatically after a power failure.”

Apple says RDMA over Thunderbolt was “co-developed and optimized for MLX Distributed and the jaccl distributed compute backend.” The macOS 26.2 release notes also name distributed AI with MLX as a use.

Harbor may handle some of these steps for you. The listing promises “no manual configuration.” Claris has not said whether the Recovery step is still needed. The “Enable in Privacy & Security…” link in the screenshot hints at a different path, but how it works is unknown.

Harbor vs exo vs MLX distributed

exo is a free, open-source app that runs one model across several devices. An exo Mac cluster works today, with no wait. MLX distributed is MLX’s own way to run across machines. Our Mac Studio cluster guide covers its setup.

exo can also set up Thunderbolt networking for you. It installs a small helper and turns off Thunderbolt Bridge. Its README says every Mac must run the exact same macOS version. On a Mac Studio, the Thunderbolt 5 port next to Ethernet can’t be used for RDMA.

Who each one fits today:

• Choose exo if you want a Mac cluster now. It ships today. Its API matches the OpenAI, Claude and Ollama formats that many apps already use.
• Wait for Harbor if you want Apple’s own app with no Terminal steps and can live without a date or price.
• Use MLX distributed if you write code and want full control.
• If you aren’t set on Macs, our DGX Spark cluster guide covers linking two NVIDIA DGX Sparks. For the wider choice, see Mac vs PC for local LLMs.

For a single Mac, LM Studio is a popular app that MacRumors mentions alongside exo.

FAQ

Is Claris Harbor real?

Yes. Claris, an Apple company, has an early-access page for it, marked “Private Preview.” Neither Apple nor Claris has put out a press release.

Can I download Harbor?

No. An App Store listing appeared on November 10, 2026 and was pulled the same day. The only official step is the Claris sign-up form.

Is Harbor free? How much does it cost?

Unknown. Claris hasn’t named a price. The page says there are “zero cloud fees” and “no per-token billing.” That is about running models locally. It says nothing about the price of the app.

Does Harbor work on an M1, M2 or M3 Mac?

Unknown for a single Mac. Aaron Perris opened the app on an M2 Pro MacBook Pro but didn’t test it. For clustering, Claris says Macs connect over Thunderbolt 5, and base M1, M2 and M3 Macs don’t have it.

Do I need Thunderbolt 5 cables?

Claris says to connect machines “over Thunderbolt 5.” Apple’s TN3205 describes Thunderbolt cables between Thunderbolt 5 ports. Neither Claris nor Apple lists a specific cable for Harbor.

Which models can Harbor run?

Open-weight models from Hugging Face, per the listing. The screenshots show gpt-oss-20b in an MLX build. Claris hasn’t listed supported models or file formats.

Is Harbor the same as exo?

No. exo is a free, open-source project from another team, and it ships today. Both pool memory across Macs over Thunderbolt. See Harbor vs exo vs MLX distributed.

Does my data leave my Mac?

Claris says “Data and prompts stay strictly within your physical Mac hardware.” We can’t test that yet. The app does download models from Hugging Face, per the screenshots.

What we’ll update when Harbor ships