Own AI agents on a Mac, offline: Osaurus is trending on GitHub

The open-source app Osaurus has entered GitHub Trending among Swift projects. It runs AI agents directly on a Mac and works with any model, local or cloud.

The project calls itself a “harness” for AI: a layer between the person and the model. The model only answers requests, while the harness keeps memory of past work, connects tools, lets agents act on their own and run code. All of that memory lives on the Mac itself, not on a third-party service’s servers. Each agent has a cryptographic identity: a digital credential that shows who exactly is acting.

With local models the app works fully offline. Cloud providers can be connected when more power is needed, and nothing leaves the Mac unless the user chooses. The app is written in Swift for Apple Silicon, meaning Macs with M-series chips, and is built without Electron (a browser-engine shell that many desktop apps are made on). The MIT license means the code is open and can be used and modified for free. Releases ship in two builds: a standard one and one with Raptor 0.6 bundled, about 4 GB.

Models can be run on a Mac without Osaurus: Ollama, which is open source, and LM Studio do that. They mainly run the model itself and give it a chat or an API. Osaurus bets on everything around it: memory, tools and agents, with the model swappable. The authors’ logic is simple: models get cheaper and replace one another, while personal context and memory cannot be replaced.

The repository was created on August 17, 2025. It now has 8,138 stars and 491 forks, copies of the project that developers take for their own changes.