Apple is building a new macOS application called Harbor that converts Macs into private AI inference servers, allowing users to run open-weight models entirely on-device. Leaks from Aaron Perris suggest models won’t need cloud connections—data and compute stay local. Multiple Macs can combine their resources via Thunderbolt 5 to tackle larger or more complex AI workloads. Harbor is now in private preview under Apple’s Claris brand, with a public release expected in the coming months.
Harbor is aimed at simplifying AI model deployment for Mac users. Once installed, the app auto-detects compatible Macs linked via Thunderbolt 5 and forms clusters without manual fiddling or command line work. Through its built-in model browser, users will be able to search, download, and load open-weight AI models from within Harbor itself—eliminating external setup hassles or infrastructure concerns.
Cluster scaling & privacy built-in
The standout feature of Harbor is its clustering ability. Users can connect two or more Macs with Thunderbolt 5 ports to share unified memory pools. This lets them run AI models that would exceed a single machine’s capacity or serve more simultaneous requests. The key here is maintaining privacy: everything happens locally, and no data—or any part of the AI model workflow—will touch the cloud.
Where this fits into Apple’s AI trajectory
The move reflects Apple’s heightening focus on on-device AI over cloud-based alternatives. With hardware like the Mac mini, Mac Studio, and now more affordable models such as the MacBook Neo, Apple has steadily ramped up its capacity for running sophisticated machine learning workloads locally. Harbor appears to formalize that direction—offering a polished user interface and workflow for model management and serving.
At present, Harbor is in restricted testing, meaning many of its features are subject to change. Public availability hasn’t been locked in, but sources indicate it could arrive in weeks or months. Meanwhile, the requirements—most notably Thunderbolt 5—signal that this won’t work on older Macs or those lacking that hardware link.
Why this matters: Harbor could reshape how creatives, developers, researchers, and privacy-minded users work with AI. By removing cloud dependencies, it addresses concerns about data security, latency, and cost. It also underscores Apple’s strategy of merging its silicon advances with software tools that amplify on-device processing. Going forward, seeing which models are supported, how efficient clustering proves in real-world settings, and whether there are limits to model size or compatibility will be key.
