Osaurus, an open-source macOS app for running AI agents against local models, says it has passed 7,300 GitHub stars and 185,000 downloads after its latest launch. The project also finished as the No. 2 Product of the Day on Product Hunt, according to the company.
The pitch is aimed squarely at Mac users who want agentic workflows without piping their files, browser state and calendar data through a cloud assistant by default. Osaurus is built in Swift and uses the Mac as the primary execution environment for open models, with hosted models available through the same interface when users choose to use them.
That puts it in the growing bucket of desktop AI tooling trying to turn local inference from a hobbyist ritual into something closer to normal software. Mac users running local models often bounce between model managers, runtimes and scripts. Agent software adds another problem: the useful version needs access to sensitive tools and context.
Osaurus handles that by acting as a control layer for models and agents on macOS. The company says each agent runs inside a hardware sandbox that limits access to memory, files and tools under user control. The app also works as a Model Context Protocol server, with more than 20 native plugins covering mail, calendar, browser access, vision, version control and the filesystem.
MCP matters because it gives models a standard way to call tools and read external context. In practice, that is the difference between a chatbot that answers from a prompt window and an agent that can act on local state. It is also where the risk lives, since the agent needs permissions to do anything useful.
The project’s local-first design is paired with interoperability rather than a walled garden. Osaurus can use model directories already shared with LM Studio, Ollama and MLX, so users do not have to download duplicate copies of the same weights to test it. The company also says its custom quantization lets larger models run at smaller quantization levels than users could otherwise fit on their machines.
The local-first macOS AI agent platform has added persistent memory, autonomous execution, cryptographic identity and voice control as community requests have grown, according to Osaurus. The project is MIT-licensed and runs on modern Mac hardware.
Terence Pae, co-founder of Osaurus, said users want modern AI capabilities without turning over files, memory and identity to a third-party cloud. He said local models are the default foundation for the app, allowing agents to work with files, browsers and system context on the user’s Mac.
The company says the download and GitHub numbers came without paid marketing. That is a useful caveat, though stars and Product Hunt rankings are soft signals. The more consequential test is whether tools like Osaurus can make local agents reliable enough for regular work while keeping the sensitive context on the machine that already holds it.
