An open-source, native macOS AI harness that sits between you and any model—local (MLX, Apple Foundation) or cloud (OpenAI, Anthropic, Gemini)—while keeping your memory, files, tools, and identity on your Mac. Unlike model runners that focus on inference, Osaurus focuses on the irreplaceable layer: agents that remember context, execute code in an isolated Sandbox (a Linux VM on macOS 26+), and stay reachable via encrypted relay links. Its Privacy Filter (experimental) scans and redacts sensitive data on-device before anything is sent to a cloud provider. Every participant gets a cryptographic identity with scoped access keys. Hardware caveat: running local models requires 64GB RAM minimum, with 128GB recommended for larger models like DeepSeek V4. Platform: Mac (native Swift, Apple Silicon) is the primary platform; Windows 11 beta launched September 2026 with a subset of features (no persistent memory, automations, or sandbox yet). Pricing: Core app is free and open source (MIT). Osaurus Cloud uses credits ($1 = 10,000) for cloud models and hosted media generation, with a free welcome credit. You can also bring your own API keys. 112K+ downloads, ~6.8K GitHub stars, founded by a former Tesla/Netflix engineer.

Pros & Cons:

Pros: The “harness” positioning is genuinely different from model runners—it treats memory, tools, and identity as the durable layer while models remain swappable, which is the right architecture for long-term personal AI use; security is designed in, not bolted on: a hardware-isolated Sandbox for code execution, an on-device Privacy Filter that redacts data before cloud calls, and per-agent cryptographic access keys with revocation; it is fully open source (MIT) and native Swift on Apple Silicon, with no Electron overhead; the MCP server implementation and 20+ native plugins (Mail, Calendar, Browser, Git, Filesystem, etc.) make it a capable integration hub for both local and cloud workflows.

Cons: The hardware bar is steep—64GB RAM minimum for local models, 128GB recommended for larger ones, which excludes most Macs and makes the “local” value proposition expensive to access; Windows support is beta and incomplete—persistent memory, automations, and sandboxing are still missing, so PC users do not get the full harness experience yet; the product is complex—Sandbox vs Working Folder, Orchestrator vs subagents, Privacy Filter, Identity, Relay, Schedules, Watchers—understanding which mode to use requires documentation study; independent third-party validation beyond founder interviews and GitHub activity is limited—no G2/Capterra/Trustpilot reviews surfaced; local performance depends heavily on Apple Silicon—on older or lower-RAM Macs, cloud models will be the only practical option.

Who is it for?:Mac users with 64GB+ RAM who want a local-first AI agent that can remember, run code safely, and access files without sending data to the cloud; privacy-sensitive individuals or teams (legal, healthcare) that need on-device redaction and cryptographic identity controls; developers who want an MCP server with native plugins and a sandbox for agentic workflows; users who want to switch freely between local models and cloud providers without losing context. Not suited for users with Macs under 64GB RAM (local models will not run well), Windows users needing the full feature set today (beta is limited), anyone seeking a simple GUI model runner (LM Studio is simpler), or teams requiring enterprise support/SLAs (community-driven, early-stage).