The news
Claris, the Apple subsidiary best known for FileMaker, is testing an application called Harbor that turns a single Apple silicon Mac into a private AI inference server. The app also supports linking multiple Macs over Thunderbolt 5 to combine their unified memory for larger models. Harbor lets users discover connected devices, browse open-weight models, download them, and load them without manual setup. The software keeps all inference on local hardware.
Context
Prior local-AI tools on macOS have required command-line configuration, separate runtimes, or cloud fallbacks. Harbor removes those steps by presenting a straightforward interface that detects Thunderbolt 5 links automatically. The feature relies on RDMA support added in macOS Tahoe 26.2, which enables low-latency memory access across clustered Macs. The testing program is limited to Claris employees and select external participants at present.
How Harbor works
Harbor runs on any Apple silicon Mac and can operate standalone or as part of a Thunderbolt 5 cluster. When multiple machines are connected, the app aggregates their unified memory pools so that models exceeding the capacity of one device can still load. The RDMA capability introduced in macOS Tahoe 26.2 reduces the overhead of moving model weights and activations between nodes. Users interact with the app through a simple browser-style interface to search for, download, and run open-weight models from public repositories. No port forwarding, firewall rules, or container configuration is required.
The source summary states that data never leaves the local hardware and that nothing is sent to the cloud. Harbor detects devices connected over Thunderbolt 5 and presents them in the interface so users can choose whether to run a model on one machine or spread it across the cluster. The emphasis is on keeping the entire workflow inside the user's own hardware without external dependencies.
Limitations of the current test
The program remains closed. Only Claris staff and a small group of external testers have access, so the exact hardware requirements, model size limits, and performance numbers are not yet public. The app currently focuses on open-weight models rather than proprietary or fine-tuned weights that users might bring from elsewhere. Thunderbolt 5 connectivity is required for clustering; older Thunderbolt versions or Wi-Fi links are not supported for memory pooling.
Why it matters
For developers and technical users who need to keep prompts and model outputs inside their own network, Harbor removes the friction that has kept most local inference tools in the hands of specialists. The ability to pool memory across Thunderbolt 5 links gives a practical path to running models that would otherwise demand a single high-end workstation or external GPU rig. Because the app is being developed inside Claris rather than inside the main Apple software organization, it may remain a niche offering rather than a mainstream macOS feature. If the current testing leads to a public release, the main change will be easier access to private inference for teams already invested in Apple silicon hardware.
The concrete outcome is that organizations with several Apple silicon Macs can now treat them as a pooled compute resource without buying new servers or exposing data to third-party clouds. Whether that stays limited to Claris customers or expands further will depend on what Apple decides after the test.
---
Sources:
No comments yet