About this project

Osaurus is a native macOS application that acts as a local harness for AI agents, written in Swift for Apple Silicon and released under the MIT license. It positions itself between the user and any model, local or cloud, and keeps context, memory, tools, and identity on the user's machine. Core capabilities described in the README: - Agents: each agent has its own prompts, memory, and theme. Tools and skills are selected automatically via RAG search. Custom agents can run an agent loop with a chosen working folder, gaining file, search, and git tools, plus optional sandboxed shell access. A built-in Orchestrator answers questions about the app, plans setting changes behind approval cards, and delegates tasks to subagents. - Sandbox: agents execute code in an isolated Linux VM using Apple's Containerization framework, with per-agent Linux users and a vsock bridge back to the host. On older macOS versions it falls back to a native Seatbelt sandbox. - Memory: three layers (identity, pinned facts, per-session episodes) plus transcript fallback, with salience scoring and background consolidation. - Privacy Filter: an on-device classifier plus regex detects names, emails, phones, URLs, addresses, dates, account numbers, and secrets before sending to cloud models, replacing them with placeholders and unscrubbing streamed replies. Described as fail-closed. - Identity and Secure Channel: secp256k1 addresses for participants, authority chains from a master key, portable access keys, and end-to-end encrypted agent-to-agent communication using X25519 and ChaCha20-Poly1305. - Relay: exposes agents via WebSocket tunnels with per-agent URLs. - Models: local MLX inference for families such as Gemma, Qwen, GPT-OSS, and Llama; Liquid AI LFM support; Apple Foundation Models on macOS 26+; and cloud providers including OpenAI, Anthropic, Gemini, xAI/Grok, Venice AI, OpenRouter, Ollama, LM Studio, and Osaurus Router. - MCP: Osaurus is both an MCP server (via `osaurus mcp` stdio bridge) and an MCP client that can aggregate remote URL-based MCP providers with OAuth 2.1. - Tools and plugins: a CLI for installing, listing, creating, and developing plugins; built-in Apple app tools (Calendar, Reminders, Contacts, Notes, Mail, Messages, Maps, Music, Shortcuts); native plugins for Vision, XLSX, PPTX, Images, Telegram, Resend, Fetch, and more. - Additional features: skills and methods importable from GitHub, scheduled automation and folder watchers, on-device voice transcription, App Intents for Shortcuts/Spotlight/Siri, browser use with per-agent WebKit sessions, offline image generation, and developer tools for server exploration and inference monitoring. Compatibility: OpenAI, Anthropic, and Ollama compatible endpoints on localhost port 1337, with streaming tool call deltas. Installation is via Homebrew cask or a DMG, requiring macOS 15.5+ and Apple Silicon. Telemetry uses Aptabase for anonymous usage analytics (off in source builds) and Sentry for crash reporting (opt-out).