Your Mac becomes the persistent brain for any LLM you plug in

What it does

Osaurus is a native Swift harness for macOS that sits between you and any language model—local or cloud—and gives AI agents persistent memory, tool use, and cryptographic identity without shipping your data off the machine. It runs fully offline with local MLX models, but can dial out to OpenAI, Anthropic, or others when you need heavier firepower. Agents execute code inside an isolated Linux VM via Apple’s Containerization framework, and a built-in privacy filter scrubs PII before it reaches cloud APIs.

The interesting bit

The privacy architecture is unusually paranoid in a good way: an on-device sparse-MoE classifier detects sensitive data, replaces it with stable placeholders like [PERSON_1], and fails closed if anything leaks through—plus you can inspect the exact wire-level bytes in an Insights panel. Memory is also aggressively compressed: conversations are distilled once at session end, scored by salience, and pruned so most turns inject fewer than 800 tokens, preventing the usual context-window bloat.

Key highlights

• Model-agnostic harness: swap between local MLX models, Apple Foundation Models, Liquid AI LFMs, or cloud providers without losing agents or memory.
• Full MCP server and client: exposes your tools to any MCP-compatible client, or aggregates remote tools from ~25 SaaS providers via HTTP/SSE.
• Cryptographic identity chain: every human and agent gets a secp256k1 address rooted in your iCloud Keychain, with revocable portable access keys.
• Native plugin ecosystem with hot-reload dev mode, plus 20+ built-in plugins for Mail, Calendar, Browser, Git, and macOS system integration.
• On-device voice transcription via FluidAudio on the Apple Neural Engine; no audio leaves the Mac.

Caveats

• Requires macOS 15.5+ and Apple Silicon, so Intel Macs and Linux/Windows users are out of luck.
• The sandboxed Linux VM requires macOS 26 (Tahoe), meaning the safest code execution path is currently locked behind an unreleased OS version.
• The privacy filter relies on a ~2.8 GB local MLX model, which is a non-trivial disk cost for a guardrail.

Verdict

Mac developers who want to own their AI stack—keeping memory, tools, and identity local while still using cloud models when convenient—should take a look. If you don’t live in the Apple ecosystem, this is not your harness.