Osaurus

Osaurus sits between you and any model, local or cloud, and provides the layer that makes AI personal: agents that remember across sessions, execute autonomously, run real code in a sandbox, and stay reachable from anywhere. Models come and go — the harness around them is what compounds.

Why it makes the list

Most AI apps keep your context on someone else's server. Osaurus is built the other way: it's a native Swift app for Apple Silicon (no Electron) that works fully offline with local MLX models, and only talks to a cloud provider when you explicitly ask it to. It's MIT licensed, with the harness — agents, memory, tools, identity — designed to outlast any one model.

Standout features

• Custom agents with their own prompts, memory, and visual theme — tools and skills are auto-selected via RAG, no manual wiring

• Sandboxed execution in an isolated Linux VM (or a native Seatbelt sandbox pre-Tahoe) with a full dev environment — shell, Python, Node.js, compilers

• Layered memory — identity, pinned facts, and per-session episodes — that decays and consolidates instead of bloating every prompt

• On-device privacy filter that redacts names, emails, secrets, and keys before anything reaches a cloud model, and fails closed if a leak slips through

• Cryptographic identity — every human, agent, and device gets its own secp256k1 address, with scoped, revocable access keys

• OpenAI, Anthropic, Ollama, and MCP compatible, so it drops into tooling you already use

Pricing

Free and MIT licensed, forever — no account required, and running local models has no usage charges. An optional Workspaces plan (from $20/workspace/month) adds shared agents and a team credit pool for cloud models.