Osaurus, an open-source Mac-only LLM server, has crossed 112,000 downloads in under a year and now operates as a routing layer between local AI models running on Apple silicon and cloud providers including OpenAI, Anthropic, and Gemini. The project lets users pick the model best suited for a given task while keeping memory, files, and tool access pinned to the user's own hardware. Co-founder Terence Pae, a former Tesla and Netflix engineer, is pitching it as a consumer-friendly alternative to developer-leaning harnesses like OpenClaw and Hermes.
The hardware bar is steep. Running models locally requires at least 64 GB of RAM, and Pae recommends 128 GB for larger systems like DeepSeek V4. That cuts the addressable market to higher-spec Mac Studio, Mac Pro, and recent MacBook Pro configurations, which is also roughly where Apple's unified-memory architecture starts to make local inference economically interesting.
Osaurus grew out of Dinoki, a desktop AI companion Pae describes as an "AI-powered Clippy." Dinoki users kept asking why they should pay for a desktop app on top of per-token cloud fees. "That's how Osaurus started," Pae said. "You can do pretty much everything on your Mac locally, like browsing your files, accessing your browser, accessing your system configurations. I figured this would be a great way to position Osaurus as a personal AI for individuals."
Key facts
- 01Osaurus has surpassed 112,000 downloads since launching nearly a year ago as an open-source, Apple-only LLM server.
- 02The app requires at least 64 GB of RAM for local models and 128 GB for larger ones like DeepSeek V4.
- 03It ships with more than 20 native plugins covering Mail, Calendar, Vision, Browser, Git, Filesystem, and Search.
- 04Osaurus routes to local models including MiniMax M2.5, Gemma 4, Qwen3.6, GPT-OSS, Llama, and to cloud providers including OpenAI, Anthropic, and Gemini.
- 05Co-founders Terence Pae and Sam Yoo are going through the New York-based Alliance accelerator.
The architecture is a harness: a single interface that brokers requests across model backends and exposes a consistent tool layer to whichever model the user selects. On the local side, Osaurus supports MiniMax M2.5, Gemma 4, Qwen3.6, GPT-OSS, Llama, DeepSeek V4, Apple's on-device foundation models, and Liquid AI's LFM family. On the cloud side, it connects to OpenAI, Anthropic, Gemini, xAI's Grok, Venice AI, OpenRouter, Ollama, and LM Studio.
“Last year, local AI could barely finish sentences, but today it can actually run tools, write code, access your browser, and order stuff from Amazon.”— Jaeden Schafer
Osaurus also functions as a full Model Context Protocol server, meaning any MCP-compatible client can reach into its tool catalog. It ships with more than 20 native plugins spanning Mail, Calendar, Vision, macOS Use, XLSX, PPTX, Browser, Music, Git, Filesystem, Search, and Fetch. A recent update added voice capabilities.
Security is the usual sticking point for tools that grant an AI live access to files, browsers, and system settings. Osaurus runs operations inside a hardware-isolated virtual sandbox that limits the model's scope on the host machine. That is a sharper boundary than OpenClaw-style developer harnesses, which Pae argues have left holes in past releases.
Pae thinks local inference is on a steep enough efficiency curve to keep eating into cloud workloads. "I can see the potential of it, because the intelligence per wattage — which is like the metric for local AI — has been going up significantly," he said. "It's on its own curve of innovation. Last year, local AI could barely finish sentences, but today it can actually run tools, write code, access your browser, and order stuff from Amazon [...] it's just getting better and better."
The team, which includes co-founder Sam Yoo, is in the New York-based Alliance accelerator and is sizing up commercial paths into regulated verticals such as legal and healthcare, where on-premise inference helps with privacy and audit requirements. Pae's argument is that an on-prem Mac Studio fleet can carry meaningful workload without a data-center dependency. "Instead of relying on the cloud, they can actually deploy a Mac Studio on-prem, and it should use substantially less power," he said. "You still have the capabilities of the cloud, but you will not be dependent on a data center to be able to run that AI."
The bet is not without friction. Local models still trail frontier cloud systems on the hardest reasoning tasks, the RAM requirements lock out most existing Mac owners, and any harness that brokers between providers will face pressure as those providers race to keep users inside their own native apps. Osaurus also has to keep pace with the plugin and tool ecosystems being built directly into Claude Desktop, ChatGPT, and Gemini.
The interesting wedge for Osaurus is that it is not trying to beat OpenAI or Anthropic on model quality — it is trying to be the place where a user keeps their files, memory, and tool permissions while renting intelligence from whichever vendor is currently ahead. That model-agnostic stance gets more valuable as the frontier shuffles every few months, and it puts Osaurus in the same conceptual bucket as routers like OpenRouter, but with the consumer surface area on the Mac itself. If local-inference economics keep improving at the pace Pae describes, the harness layer — not the model — may end up being the part of the stack consumers actually pay for.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.




