Osaurus
A tiernew this weekA native, open-source macOS harness that lets you swap local and cloud AI models freely while keeping your memory, history, and tools locked on your own hardware.
Kai's verdict
Osaurus is the most serious attempt yet at a truly native macOS AI harness — the MCP integration, local memory, and sandboxed code execution put it well ahead of Ollama's raw-server approach, and the privacy story is genuinely compelling for the right user; the catch is that you need serious Apple Silicon hardware to get the most out of it. (Verdict pending Phi's full review.)
Strengths
- True hybrid inference — seamlessly switch between local MLX models and cloud providers (OpenAI, Anthropic, Gemini, Grok, OpenRouter, etc.) within a single persistent session
- Full MCP server with 20+ native plugins (Mail, Calendar, Browser, Git, Filesystem, Music, XLSX and more), making it a genuine macOS automation layer
- Four-layer local memory system (profile, working memory, summaries, knowledge graph) that compounds context over time — entirely on-device
- Sandboxed Linux VM for code execution via Apple's Containerization framework, so agents can run real shell/Python/Node code without touching your Mac
- Pure Swift, ~10 MB, no Electron — near-instant startup and far lower resource usage than most competing GUIs
Weaknesses
- Apple Silicon and macOS 15.5+ exclusive — zero support for Windows, Linux, or Intel Macs
- Serious local models (7B+) demand 32–128 GB of RAM, pricing out most non-pro Mac users from the flagship local-first experience
- Still early-stage and founder-led; plugin ecosystem and long-term support are unproven at scale
Best for
Privacy-conscious Mac power users and developers on Apple Silicon who want a single, always-on AI runtime that blends local models with cloud APIs — without surrendering their context to any vendor's servers.
Pricing
Free (open source, MIT license)
Completely free with no usage limits; you pay only for any cloud API keys you choose to bring (e.g. OpenAI, Anthropic).