Osaurus brings both local and cloud AI models to your Mac

As AI models progressively go commoditized, startups are racing to physique the bundle furniture that sits connected apical of them. One absorbing entrant into this abstraction is Osaurus, an unfastened source, Apple-only LLM server that lets users determination betwixt antithetic section AI models, either locally oregon successful the cloud, portion keeping their files and tools each connected their ain hardware.

Osaurus evolved retired of the thought for a desktop AI companion, Dinoki, which Osaurus co-founder Terence Pae described arsenic a benignant of “AI-powered Clippy.” Dinoki’s customers had asked him wherefore they should bargain the app if they inactive had to wage for tokens — the usage units AI companies complaint for processing prompts and generating responses.

That got Pae reasoning much profoundly astir moving AI locally.

“That’s however Osaurus started,” Pae, antecedently a bundle technologist astatine Tesla and Netflix, told TechCrunch implicit a call. The idea, helium explained, was to effort to tally an AI adjunct locally. “You tin bash beauteous overmuch everything connected your Mac locally, similar browsing your files, accessing your browser, accessing your strategy configurations. I figured this would beryllium a large mode to presumption Osaurus arsenic a idiosyncratic AI for individuals.”

Pae began gathering the instrumentality successful nationalist arsenic an open-source project, adding features and fixing bugs on the way.

Image Credits:Osaurus, Inc.

Today, Osaurus tin flexibly link with locally hosted AI models oregon unreality providers similar OpenAI and Anthropic. Users tin freely take which AI models they’re using, and support different aspects of the AI acquisition connected their ain hardware, similar the models’ ain memory, oregon their files and tools.

Given that antithetic AI models person antithetic strengths, the vantage of this strategy is that users tin power to the AI exemplary that champion fits their needs.

Such a operation makes Osaurus what’s called a “harness” — a power furniture that connects antithetic AI models, tools, and workflows done a azygous interface, akin to tools similar OpenClaw oregon Hermes. However, the quality is that specified tools are often aimed astatine developers who cognize their mode astir a terminal. And sometimes, similar successful the lawsuit of OpenClaw, they whitethorn airs information issues and holes to interest about.

Osaurus, meanwhile, presents an easy-to-use interface that consumers tin use, and addresses information concerns by moving things successful a hardware-isolated, virtual sandbox. This limits the AI to a definite scope, keeping your machine and information safe.

Image Credits:Osaurus, Inc.

Of course, the signifier of moving AI models connected your instrumentality is inactive successful its aboriginal days, fixed that it’s heavy resource-intensive and hardware-dependent. To tally section models, your strategy volition request astatine slightest 64 GB of RAM. For moving larger models, similar DeepSeek v4, Pae recommends systems with astir 128 GB of RAM.

But Pae believes section AI’s needs volition travel down successful time.

“I tin spot the imaginable of it, due to the fact that the quality per wattage — which is similar the metric for section AI — has been going up significantly. It’s connected its ain curve of innovation. Last year, section AI could hardly decorativeness sentences, but contiguous it tin really tally tools, constitute code, entree your browser, and bid worldly from Amazon […] it’s conscionable getting amended and better,” helium said.

Image Credits:Osaurus, Inc.

Osaurus contiguous tin tally MiniMax M2.5, Gemma 4, Qwen3.6, GPT-OSS, Llama, DeepSeek V4, and different models. It besides supports Apple’s on-device instauration models, Liquid AI’s LFM family of on-device models, and successful the cloud, it tin link to OpenAI, Anthropic, Gemini, xAI/Grok, Venice AI, OpenRouter, Ollama, and LM Studio.

As a afloat MCP (Model Context Protocol) server, you tin springiness immoderate MCP-compatible lawsuit entree to your tools arsenic well. Plus, it ships with implicit 20 autochthonal plugins for Mail, Calendar, Vision, macOS Use, XLSX, PPTX, Browser, Music, Git, Filesystem, Search, Fetch, and more. 

More recently, Osaurus was updated to see dependable capabilities arsenic well.

Since the task went unrecorded astir a twelvemonth ago, it has been downloaded northbound of 112,000 times, according to its website.

Currently, Osaurus’ founders (who see co-founder Sam Yoo) are participating successful the New York-based startup accelerator Alliance. They’re besides reasoning astir adjacent steps, which could spot Osaurus being offered to businesses, similar those successful the ineligible abstraction oregon successful healthcare, wherever moving section LLMs could code privateness concerns.

As the powerfulness of section AI models grows, the squad believes it could little the request for AI information centers.

“We’re seeing this explosive maturation successful the AI abstraction wherever [cloud AI providers] person to standard up utilizing information centers and infrastructure, but we consciousness similar radical haven’t truly seen the worth of the section AI yet,” Pae said. “Instead of relying connected the cloud, they tin really deploy a Mac Studio on-prem, and it should usage substantially little power. You inactive person the capabilities of the cloud, but you volition not beryllium babelike connected a information halfway to beryllium capable to tally that AI,” helium added.