Osaurus is a macOS native artificial intelligence agent execution and management runtime harness, fully optimized for Apple Silicon hardware. Just as a smartphone's operating system acts as a central control unit that runs various mobile apps and precisely allocates device resources, Osaurus serves as a control center and execution engine, enabling local and cloud large language models to safely manipulate the user's Mac system files and development tools. It is released as a complete open-source project under the MIT license, meticulously excluding the commonly used, heavy Electron wrapper and compiled solely in pure Swift, harmonizing with Apple's latest graphics and computation APIs. This minimizes the Mac's hardware resource usage and provides extremely responsive user interaction. Existing simple model execution clients have demonstrated clear limitations in safely controlling a practical working environment, such as allowing the agent to directly write code and execute system-internal commands, relying solely on the text inference performance of large language models. To address this technical need, Osaurus introduces a unique Linux virtual machine (VM)-based sandbox architecture leveraging Apple's Containerization Framework. This creates a complete agent loop where the agent repeatedly installs packages, compiles, and executes scripts within a virtual environment, completely isolated to prevent any harm to the host macOS system. Furthermore, it presents a robust offline security pipeline through an on-device privacy filter that operates within the device, detecting and masking sensitive information such as names, emails, phone numbers, system passwords, and API tokens locally before transmitting them to the server, even when using external commercial cloud models. For researchers and developers in various domains, including computer science and life sciences, Osaurus is a useful tool that tirelessly and autonomously performs the roles of a powerful offline assistant and secretary. By fully adopting the Model Context Protocol (MCP), it allows users to instantly combine over 20 external productivity service APIs, such as Notion, Linear, GitHub, and Stripe, with just a few mouse clicks. It can also be combined with macOS native plugins, such as email, calendar, and document viewers, to independently plan and execute complex multi-step tasks. Additionally, it features a Watcher function that monitors specific folders in the local repository in real-time, along with a three-tiered intelligent RAG knowledge retrieval memory layer divided into Session Memory and Pinned Facts. This ensures that whenever a new file or paper is detected, only the core information is extracted and accumulated in long-term memory without unnecessary token waste, allowing the user to complete designated automated tasks perfectly even while they are asleep or away.

Osaurus is a macOS native artificial intelligence agent execution and management runtime harness, fully optimized for Apple Silicon hardware. Just as a smartphone's operating system acts as a central control unit that runs various mobile apps and precisely allocates device resources, Osaurus serves as a control center and execution engine, enabling local and cloud large language models to safely interact with the user's Mac system files and development tools. It is released as a fully open-source project under the MIT license and is commonly used.