Apple Tests Harbor for Local AI Model Deployment on Mac Computers

Apple subsidiary Claris, known for developing the FileMaker database platform, is testing the Harbor application for locally running artificial intelligence (AI) models on Mac computers. The program is designed to work with models with open weights directly on devices with Apple Silicon processors, without the need to send requests to external cloud servers.

Information about Harbor emerged on November 10, 2026, following software researcher Aaron Perris's publication of a description of the application, previously known as Fabric. The program is in closed preliminary testing. Its page was available in the App Store via a direct link but later disappeared from the store.

Harbor will allow combining multiple Macs for AI work

According to the published description, Harbor can run AI models on a single computer or multiple devices connected via Thunderbolt 5, a high-speed data transfer interface. The application is supposed to automatically detect connected Macs and combine their computing resources, allowing the use of aggregate memory for working with larger models.

This approach is intended for running models that lack the memory of a single computer, as well as for processing a larger number of requests simultaneously. Users will not need to manually configure the connection via the command line.

The application also provides for finding and downloading models directly through the interface. Published demo images show work with gpt-oss-20b, an OpenAI model with open weights hosted in the Hugging Face catalog.

According to Claris, with local processing, requests and user data remain on the user's own equipment. This format allows avoiding payments to cloud providers for each processed token. However, computers with sufficient computing resources are required, and the cost of Harbor itself is still unknown.

Apple application enters the market with existing alternatives

Local AI model execution on Mac is already supported by other software solutions. For example, the free tool Exo allows combining multiple devices for distributed computing, while LM Studio provides the ability to run models on a single computer.

The main difference of Harbor, according to its description, lies in combining these functions in one interface with automatic device setup and model management. However, independent test results that would allow comparing Harbor's performance with competitors have not yet been presented.

Claris has not announced the cost of the application or the terms of its distribution. The registration form for preliminary testing requests the company name, which may indicate the project's focus on corporate users.

The public release date of Harbor, the full list of supported models, and the final system requirements remain unknown.