Osaurus is a native macOS application designed to serve as a harness for AI agents. It empowers users to run any AI model entirely on their local machine, providing features such as persistent memory, autonomous execution capabilities, and cryptographic identity. Built with Swift for Apple Silicon, Osaurus prioritizes user privacy and data security by operating fully offline and remaining open source.
The core problem Osaurus addresses is the prevalent practice of AI assistants running on third-party servers. This often means user files, memory, and context are transmitted off the user's machine, frequently behind a subscription model. Osaurus offers an alternative by keeping all data and processing local, giving users full control and ownership over their AI interactions and data.
Key features of Osaurus include the ability for agents to interact with local macOS resources such as Calendar, Contacts, and iMessages, as well as work with the user's real files. It provides an isolated sandbox environment for agents to write and run code, generating outputs like images, PDFs, and presentations directly on the user's desktop. A crucial aspect of its design is an approval gate for every agent action, allowing users to review and authorize what the agent intends to do before execution. For running AI models, Osaurus supports local execution fully offline, utilizing its own Swift MLX runtime without requiring Python. It also allows users to bring their own API keys for frontier models from providers like OpenAI, Anthropic, and Gemini, or opt for pay-as-you-go Osaurus Cloud, all while ensuring that user memory and context remain local.
Osaurus also supports the concept of subagents, enabling agents to delegate tasks to other local or remote models. This allows for more complex workflows where different models can be leveraged for specific parts of a task. While intelligent routing is still evolving, users can set up routing logic via system prompts to direct tasks to appropriate models, whether local or cloud-based.
The platform is built with privacy and security at its forefront. By running locally, it eliminates the need for accounts or subscriptions for basic functionality, and its open-source nature invites community contributions and transparency. The application is developed in Swift, specifically optimized for Apple Silicon, and avoids using Electron for a more native and efficient experience.
Osaurus operates by storing all data, including chat history and memory, locally on the user's file system. The Mac's own inferencing capabilities are leveraged to distill memories and process embeddings for dynamic contexts. For file operations within the sandbox, Osaurus maintains an internal log of changes, allowing agents to undo file operations, with plans to enhance this with a more visual representation.
Users benefit from enhanced privacy, data ownership, and the ability to run powerful AI agents without constant internet connectivity or recurring fees. The explicit approval gate for actions provides a significant layer of trust and control, especially when agents interact with sensitive local data or perform system-level tasks. The platform's flexibility in model selection, from local offline models to bringing your own keys, caters to a wide range of user needs and preferences.
Concrete use cases include using Osaurus as an autonomous assistant for managing a vacation rental business, handling emails, marketing, and general organization. It can also be used to introduce family and friends to AI by sharing Venice API keys. Another application is creating a personal dashboard by setting up Agent DB and self-scheduling. The platform is also being developed to support image generation and text-to-speech functionalities, potentially integrating with local or cloud server URLs for these capabilities.
Osaurus is free to use and is open source under an MIT license. It is designed for Mac users with Apple Silicon. The recommended RAM for running local models is 24GB, though 16GB can get started with smaller models. The platform is actively seeking community contributions through its GitHub repository.
In summary, Osaurus provides a powerful, private, and user-controlled environment for running AI agents locally on macOS, offering a secure and flexible alternative to cloud-based AI services.
