Osaurus Wants to Make Local AI on Mac Easier, Safer, and More Private
As artificial intelligence models become easier to access and increasingly interchangeable, a new wave of startups is focusing on what comes next: the software layer that helps people actually use those models in everyday workflows. One of the more interesting projects in this growing space is Osaurus, an open-source LLM server built specifically for Apple hardware.
Osaurus is designed to let users switch between different AI models, run them locally or through cloud providers, and keep their files, tools, and AI memory on their own machines. For anyone interested in local AI, private AI assistants, or running large language models on a Mac, Osaurus is trying to make the experience more flexible and approachable.
The project began with a different idea. Co-founder Terence Pae had been working on Dinoki, a desktop AI companion he described as an “AI-powered Clippy.” But users kept asking an important question: Why should they pay for an app if they still had to pay separately for AI tokens every time they used it?
That question pushed Pae to explore a different path: running AI locally.
The thinking was simple. A Mac can already access files, browser data, system settings, apps, and personal workflows. If an AI assistant could run on the same machine, it could become a more private and powerful personal assistant without depending entirely on cloud infrastructure.
From that idea, Osaurus began taking shape as an open-source project. Pae built it publicly, adding features, fixing bugs, and expanding its capabilities as more users began experimenting with local AI.
Today, Osaurus works as a flexible control layer for AI models. Users can connect it to local models running on their own hardware or to cloud-based services such as OpenAI, Anthropic, Gemini, xAI/Grok, Venice AI, OpenRouter, Ollama, and LM Studio. The key difference is that users are not locked into one model or one provider.
That flexibility matters because different AI models are good at different tasks. One model may be better for coding, another for reasoning, another for writing, and another for handling local tools. Osaurus lets users choose the right model for the job while keeping important parts of the AI experience, such as files, tools, and memory, under their own control.
In practical terms, Osaurus acts as a harness for AI. It connects multiple models, tools, and workflows through a single interface. While similar concepts often appeal mainly to developers comfortable with command-line tools, Osaurus is aiming for a more consumer-friendly experience. The goal is to make local AI on Mac feel less like a technical experiment and more like a usable everyday assistant.
Security is also a major part of the pitch. Giving an AI assistant access to files, browser tools, and system functions can create obvious risks. Osaurus addresses that concern by running processes inside a hardware-isolated virtual sandbox. This limits what the AI can touch and helps protect the rest of the computer and user data.
Local AI still has challenges. Running large language models directly on a personal computer is demanding, and performance depends heavily on hardware. According to Pae, users should have at least 64 GB of RAM to run local models comfortably. For larger models such as DeepSeek V4, around 128 GB of RAM is recommended.
Even so, Pae believes the hardware requirements for local AI will improve over time. He points to the rapid gains in “intelligence per watt,” a key measure for on-device AI performance. Local models that struggled to complete basic sentences not long ago can now use tools, write code, access browsers, and perform far more complex tasks.
Osaurus already supports a wide range of models, including MiniMax M2.5, Gemma 4, Qwen3.6, GPT-OSS, Llama, DeepSeek V4, and others. It also supports Apple’s on-device foundation models and Liquid AI’s LFM family of on-device models.
The platform is also a full MCP server, meaning any client compatible with the Model Context Protocol can access connected tools. Osaurus includes more than 20 native plugins, covering Mail, Calendar, Vision, macOS Use, spreadsheets, presentations, browser access, music, Git, filesystem tools, search, fetch, and more. A more recent update added voice capabilities, bringing it closer to the idea of a true personal AI assistant for Mac users.
Interest in the project has been growing. Since launching nearly a year ago, Osaurus has reportedly been downloaded more than 112,000 times.
The company’s founders, including Terence Pae and Sam Yoo, are currently participating in the New York-based startup accelerator Alliance. They are also considering future opportunities for business use, especially in industries where privacy is critical. Legal firms, healthcare organizations, and other privacy-sensitive businesses could benefit from running local LLMs on their own hardware instead of sending sensitive information to external AI servers.
That vision could become more important as demand for AI infrastructure continues to rise. Cloud AI providers are building more data centers to keep up with usage, but Osaurus is betting that local AI will become a serious alternative for some users and organizations.
Instead of depending entirely on remote infrastructure, businesses could run powerful AI systems on-premises using machines such as a Mac Studio. That approach could reduce dependence on data centers, lower power usage for certain workloads, and give users more control over privacy and security.
Osaurus is still part of an early and fast-changing local AI market, but its direction is clear. It wants to make running AI on personal hardware easier, safer, and more useful. As local models become more capable and efficient, tools like Osaurus could play a major role in how people use AI on their own devices.
