osaurus-ai/osaurus

Own your AI. The native macOS harness for AI agents -- any model, persistent memory, autonomous execution, cryptographic identity. Built in Swift. Fully offline. Open source.

What it solves

Osaurus is a local-first AI harness for macOS that allows users to own their AI context, memory, and tools without relying on cloud servers. It provides a unified interface to interact with various local and cloud models while keeping the personal data layer—identity, memory, and tool execution—on thedevice.

How it works

Built natively in Swift for Apple Silicon, Osaurus acts as a middleware layer between the user and the AI model. It manages a system of autonomous agents that can execute code in an isolated Linux VM sandbox (or a macOS Seatbelt sandbox on older versions) to ensure security. It uses a three-layer memory system (identity, pinned facts, and per-session episodes) to maintain continuity. For privacy, it includes an on-device classifier to scrub PII before sending data to cloud providers.

Who it’s for

Mac users who want a private, extensible AI assistant capable of running local models (via MLX) or cloud APIs, and who require the ability to to execute code and browse the web autonomously.

Highlights

  • Local-First Execution: Fully offline capabilities with optimized MLX inference for Apple Silicon.
  • Secure Sandboxing: Agents run code in an isolated Linux VM with a dedicated home directory and shell access.
  • Privacy Filter: On-device PII detection and redaction for cloud-model requests.
  • MCP Integration: Acts as both an MCP server (exposing local tools to other clients) and an MCP client (aggregating tools from remote providers).
  • End-to-End Encryption: Secure, encrypted communication channels between agents using X25519 and ChaCha20-Poly1305.
  • Browser Use: Dedicated per-agent WebKit sessions for autonomous web navigation and interaction.
  • Model Agnostic: Supports local models (Gemma, Llama, etc.), Liquid AI LFMs, Apple Foundation Models, and major cloud providers.