Osaurus is a free, open-source macOS app that serves as a unified layer between you and any AI model, whether local or cloud-based. Built natively in Swift for Apple Silicon, it lets you run models like Llama, Qwen, and Gemma directly on your Mac while also connecting to OpenAI, Anthropic, and other cloud providers. Each agent gets its own memory, tools, and personality. A 4-layer memory system remembers context across conversations, and a sandboxed Linux VM lets agents execute real code safely. Your data stays encrypted on your machine unless you explicitly choose otherwise.

• Hybrid Local + Cloud Inference: Run models locally on Apple Silicon via MLX or connect to OpenAI, Anthropic, Gemini, xAI, OpenRouter, Ollama, and LM Studio within a single session.
• 4-Layer Persistent Memory: Automatically extracts facts, detects contradictions, and maps relationships from every conversation so your agents remember what matters across sessions.
• Customizable AI Agents: Create purpose-built agents, each with their own prompts, memory, visual theme, and toolset. A RAG-based system auto-selects the right tools for each task.
• Sandboxed Code Execution: Toggle on a Linux VM powered by Apple's Containerization framework. Agents can run shell, Python, and Node.js code with zero risk to your Mac.
• 20+ Native Plugins: Built-in tools for Mail, Calendar, Vision, XLSX, PPTX, Browser, Music, Git, Filesystem, Search, and more, plus full MCP server support.
• On-Device Privacy Protection: Sensitive information such as names, emails, and API keys can be detected and masked locally before any request is sent to external models, helping reduce data exposure.
• Voice Input with WhisperKit: Voice-to-text input powered by WhisperKit runs locally on your Mac, allowing fast speech recognition without sending audio to the cloud.
• Remote Access via Relay: Access your local agents remotely through secure relay connections, enabling interaction with your AI system from external devices without manual network setup.

• Completely Free and Open Source: MIT licensed with no usage limits or subscription tiers. You only pay for cloud API keys if you choose to use them.
• Strong Privacy Architecture: Data is encrypted at rest, signed at every boundary, and never leaves your Mac unless you explicitly opt in. The privacy filter adds an extra layer before cloud requests.
• Native Performance: uilt in Swift and optimized for Apple Silicon, delivering fast and lightweight performance without Electron or heavy runtime overhead.
• Flexible Model Support: Switch between dozens of local models and major cloud providers within the same conversation, keeping full context and memory intact.
• Rich Plugin Ecosystem: Built-in tools and MCP support for extending workflows across files, browser, system tasks, and developer tools.
• Active Community: Open-source project with strong GitHub traction and growing adoption across the macOS AI developer ecosystem.

• macOS Only: No support for Windows or Linux. Requires Apple Silicon (M1 or newer), so Intel Macs are excluded entirely.
• High RAM Requirements: Running local models needs at least 64GB of unified memory. Larger models like DeepSeek V4 may need 128GB, putting this out of reach for many Mac users.
• Sandbox Requires macOS 26: The isolated Linux VM for safe code execution only works on macOS 26 (Tahoe) or later, limiting a key feature for users on older OS versions.
• Cloud Costs Are Separate: While the app is free, using cloud models like GPT-4 or Claude through Osaurus still requires paying those providers' API fees.

• Privacy-Sensitive Professional Work: Lawyers, healthcare professionals, and financial advisors can use AI assistance while keeping client data entirely on their local machine. The privacy filter catches sensitive information before it reaches any cloud provider, meeting strict compliance needs.
• Software Development Companion: Point an agent at your project folder and get a coding partner with git tools, file search, and the ability to run code in a sandboxed environment. The persistent memory means your agent learns your codebase conventions over time.
• Content Creation Pipeline: Writers and marketers can create agents that turn raw notes into polished drafts, convert interview transcripts into articles, or generate SEO-optimized content from product updates, all while keeping proprietary material off third-party servers.
• macOS Automation: Use the macOS Use plugin to automate interactions with any application. Set up folder watchers that trigger workflows (e.g., new invoice arrives, agent extracts data and emails a summary) or schedule recurring tasks like daily reports.
• Research and Knowledge Management: Build a personal research assistant that remembers every source, finding, and decision across sessions. The knowledge graph memory layer maps relationships between concepts, making it easy to surface relevant context months later.
• Offline AI Access: Run capable models entirely on-device with no internet connection required. Useful for travel, secure environments, or situations where you need AI assistance without depending on cloud availability.