Osaurus: A Native AI Agent Manager for macOS

Osaurus is an open-source, local-first AI harness built natively in Swift for macOS on Apple Silicon. While models are becoming interchangeable commodities, Osaurus keeps the parts that actually matter-your context, persistent memory, execution tools, and cryptographic identity-firmly anchored to your own Mac.

It acts as an intelligent intermediary between you and your models (whether running locally offline or through cloud APIs), providing the persistent runtime and security perimeter needed for autonomous agent execution.

Why You Should Consider it?

True Data Sovereignty: Run entirely offline with local models or connect to cloud providers on your own terms. Your files, memory graphs, and identity keys never touch third-party servers.

Pure Apple Silicon Performance: Written natively in Swift rather than bundled as a heavy Electron wrapper, ensuring minimal RAM overhead and swift responsiveness.

Safe, Contained Execution: Agents can run real shell scripts, compile code, and install packages without putting your host macOS operating system at risk.

Anti-Leak Cloud Safeguards: Prevents accidental data leaks to remote cloud providers via an on-device privacy filter that intercepts and scrubs sensitive personal records before transmission.

Zero Configuration Networking: Securely access your local agents from remote locations or connect agents to each other using cryptographically verified, end-to-end encrypted tunnels.

Features

• Isolated Sandbox Execution: Custom agents execute inside a sandboxed Linux VM (via Apple’s Containerization framework) with its own user space, workspace, and a secure vsock host bridge. On older macOS environments, it safely falls back to a confined Seatbelt sandbox.
• Automated Agent Loops: Assign custom agents a specific directory, and they autonomously plan tasks via markdown to-do lists, write and execute code, leverage Git, and return structured outcome summaries.
• Tri-Layer Memory System: Instead of dumping raw chat logs into prompts, Osaurus distills conversations at session end into identity profiles, pinned facts, and episodic events. It injects relevant slices via salience scoring, keeping token overhead minimal (typically ~800 tokens or less).
• On-Device Cloud Privacy Filter: Uses a local 1.5B parameter sparse-MoE classifier alongside deterministic regex to redact names, emails, API tokens, and credentials into stable tokens (e.g., [PERSON_1]), restoring them locally during streaming.
• Cryptographic Identity & Trust: Every human, agent, and device gets a secp256k1 key rooted in your master iCloud Keychain, enabling fine-grained, revocable access keys (osk-v1).
• Secure Channels & Public Relay: Agents communicate across LANs or over the web via agent.osaurus.ai WebSocket tunnels, protected with mutual X25519 handshakes and ChaCha20-Poly1305 encryption.
• The Orchestrator & Subagent Spawning: A safety-bounded built-in agent manages system configurations, queries live state, and coordinates tasks by delegating to specialized background subagents with managed memory and single-residency model swaps.

How to Install and Use

1. Install via Homebrew

brew install --cask osaurus

(Requires macOS 15.5+ and Apple Silicon hardware).

2. Launch the Application Launch it from Spotlight (Command + Space → type Osaurus) or control it via the CLI:

osaurus ui       # Open the native desktop chat interface
osaurus serve    # Start the background harness server
osaurus status   # Verify active agents and runtime health

3. Set Up Your Agents

• Use the default Orchestrator to control app preferences, answer setup questions, and delegate complex tasks.
• For coding or file operations, create a Custom Agent, select a target working directory, toggle the Sandbox, and assign it a task to run within its isolated container.

License

MIT License

Resources & Downloads

• Source-code & Downloads