Osaurus is a native macOS application that acts as an AI agent harness, running silently in the menu bar and providing a persistent runtime for AI agents with memory, autonomous execution, and tool access. The app is open-source under the MIT license and requires Apple Silicon.
Each agent in Osaurus maintains a three-layer memory system: a fixed identity layer, a set of pinned facts, and per-session episodic memory that consolidates automatically over time. Agents can execute code and shell commands in isolated Linux virtual machines powered by Apple’s Containerization framework, restricting the impact of autonomous actions. The app ships with over 20 native macOS plugins including Mail, Calendar, Vision, Git, Browser control, and file system access, enabling agents to interact with macOS directly without additional configuration.
Model routing is flexible: a single Osaurus installation can run local MLX-optimized models entirely on-device, connect to cloud providers such as OpenAI and Anthropic, or use Apple Foundation Models. Osaurus exposes an MCP (Model Context Protocol) server, making agents accessible to external tools and editors. A relay system publishes agents via secure WebSocket tunnels for access from other devices. Voice input with on-device transcription and wake-word activation is included.
System requirements: macOS 15.5 or later. Apple Silicon (M1 or newer) required—Intel Macs are unsupported. Sandbox execution in isolated VMs requires macOS 26 (Tahoe). Running large local language models requires 64 GB or more of unified memory. Install via brew install --cask osaurus or download .dmg from GitHub Releases.
Pricing: (Free, open-source) MIT license.
Limitations: Intel Mac users cannot run Osaurus. Sandbox and VM-based execution require macOS 26, not yet stable as of May 2026. Local model inference with large parameter counts demands hardware that most users do not have (64 GB+ RAM).
Alternatives: Ollama (free, open-source local model server, supports more platforms including Intel Mac and Linux); LM Studio (free, graphical local model runner, cross-platform); Jan (free, open-source, cross-platform local AI desktop).
Suitable for developers on Apple Silicon Macs who want a persistent, multi-agent AI runtime with direct macOS tool access and the option to keep all model inference on-device.
