Osaurus – Personal AI Agent Platform for macOS

What it is

Osaurus is a native macOS application (Apple‑silicon only) that acts as a local “harness” for large language models. It lets you run agents – specialised AI assistants – that keep their own memory, tools, and identity on your computer. The platform can talk to any LLM, whether it’s a model stored locally (via MLX, Apple Foundation Models, or Liquid AI) or a cloud service (OpenAI, Anthropic, Gemini, etc.). All the orchestration, privacy filtering, and tool‑calling happen on‑device, so no data leaves your Mac unless you explicitly choose a cloud model.

Core features (as described in the README)

Feature	What it does
Agents & Memory	Create custom agents with independent prompts, persistent memory (identity, pinned facts, session episodes) and visual themes. Memory is distilled after each session and kept compact.
Sandboxed execution	Each agent can run code in an isolated Linux VM (Alpine) using Apple’s Containerization framework, or fall back to a macOS Seatbelt sandbox on older OS versions.
Privacy filter	On‑device classifier (OpenAI privacy‑filter model) scrubs personal data before any request is sent to a cloud provider, with a fail‑closed policy.
Cryptographic identity	Agents and devices have secp256k1 addresses derived from your iCloud Keychain; access keys can be scoped, revoked, and used for end‑to‑end encrypted communication.
Secure channel & Relay	Agents can be exposed via a WebSocket tunnel (agent.osaurus.ai) with X25519 handshake and ChaCha20‑Poly1305 encryption; the relay never sees plaintext.
Orchestrator	Built‑in default agent that delegates work to custom agents or sub‑agents, manages budgets, and never directly accesses the filesystem or sandbox.
Sub‑agents	Spawn specialised agents (local or remote models) for a single task, with single‑residency hand‑off to conserve RAM.
Browser tool	Optional per‑agent WebKit‑based browser that can navigate, click, and type, with user‑approved actions and isolated cookie storage.
Image generation	Offline image creation/editing using local models (e.g., FLUX, Qwen‑Image) with results shown inline.
Model‑agnostic harness	Supports local models via MLX (Gemma, Llama, Qwen, etc.), Liquid AI’s LFM, Apple Foundation Models, and cloud APIs.
MCP server	Implements the Model Context Protocol, exposing tools via a stdio bridge for other MCP‑compatible clients.
Plugin ecosystem	osaurus tools command to install, list, create, and develop Swift plugins; many Apple‑app integrations are built‑in.
Voice input	On‑device transcription using FluidAudio on Apple’s Neural Engine; wake‑word activation and global hot‑key support.
System integration	App Intents expose “Ask Osaurus” and “Run Osaurus Agent” to Spotlight, Siri, and Shortcuts.
Telemetry & crash reporting	Anonymous usage analytics via Aptabase (opt‑in) and optional crash reporting via Sentry (opt‑out).

How to get it

• Install via Homebrew Cask: brew install --cask osaurus.
• Or download the .dmg from the GitHub releases page. Two variants exist:
  • Standard (~70 MB) – prompts you to download a model on first run.
  • Full (~4 GB) – bundles the Raptor 0.6 model for immediate offline use.
• Launch with Spotlight (⌘ Space → Osaurus) or the CLI (osaurus ui, osaurus serve, osaurus status).

Typical use‑case workflow

1. Pick a model – local (e.g., Raptor 0.6) or cloud (OpenAI, Anthropic, etc.).
2. Create an agent – give it a name, prompt, and optionally enable tools like file access, browser, or image generation.
3. Chat – the agent remembers past interactions, can call tools, and can spawn sub‑agents for specialised tasks.
4. Run code – enable the sandbox; the agent writes a markdown todo list, the VM executes the code, and the result is returned in the chat.
5. Persist & share – agents have cryptographic IDs; you can expose one via the secure relay URL for remote access.

Who it’s for

• Power users who want a private, offline AI assistant on macOS.
• Developers looking to build Swift plugins or integrate macOS apps (Calendar, Mail, etc.) with LLMs.
• Enterprises needing on‑device AI with end‑to‑end encryption and strict data‑privacy guarantees.

Limitations noted in the README

• Requires macOS 15.5+ on Apple Silicon; the full Linux VM sandbox needs macOS 26 (Tahoe).
• The offline image generation and some heavy models need several gigabytes of storage.
• Certain features (e.g., Apple Foundation Models) are only available on the newest macOS releases.

---

All information above is taken directly from the project's README; no external assumptions have been added.