An open source Mac-only LLM server called Osaurus hit TechCrunch on Friday. The pitch sounded like every other local-AI tool that's come out in the last two years, so I almost didn't bother. Then I noticed two things in the writeup. It was built in public for a year. And it's already been downloaded 112,000 times.

That's not nothing. So I installed it Friday night.

The thing that's actually different

Most local-AI apps assume you want one of two things. Either you're a developer who'll spend an evening in a terminal getting Ollama wired up to your editor, or you're a paid OpenAI customer who occasionally wonders if Llama 4 might be cheaper. Osaurus is for a third person I didn't really know existed: someone who wants both, in the same app, with a UI that doesn't make them feel stupid.

You open it. You pick a model. If the model lives on your hardware it runs there. If it lives in the cloud (Anthropic, OpenAI, whatever you've got keys for) it routes there. Your files, your memory, your tool configurations stay on the Mac in what Osaurus calls a "hardware-isolated sandbox." No per-token bills for the things that don't need to go to the cloud. No giving up on the cloud entirely for the things that do.

That's actually a useful position to take. I don't think anyone in the local-AI space has framed it quite this cleanly.

What I'm using it for

Three things, after three days.

First, transcript-style work. Anything where I'm pasting in a chat transcript or a long article and asking for a summary or restructure goes through a local model now. None of that ever needs to leave my machine, the cloud rate was burning credits unnecessarily, and the latency is honestly better.

Second, exploratory prompting. Half the time I'm playing with a prompt I don't actually need the smartest model. I need fast iteration. Running a local model for the throwaway versions and only kicking up to Claude or GPT when the prompt is doing something serious has reduced my Anthropic spend by enough this weekend that I noticed.

Third, anything with anything client-confidential. This is the part I should have been doing all along and wasn't, because the friction of running a separate local app was high enough to keep me in the cloud. Osaurus took that friction down to nearly zero. That's the part that matters.

What it's not for

The big models. DeepSeek v4 needs about 128GB of RAM. I have 32GB. So the heavy lifting still goes to the cloud, and it should. Local AI on consumer Macs is only going to be as good as the small-model frontier, and small models are improving but they're still small models.

Multi-modal stuff. I haven't tested image generation or audio yet, and I suspect the local story there is significantly worse than the text one.

Anything where you actually want the cloud's full context window. Local models have shorter context windows. You'll hit the limit on long codebases or full PDFs faster.

The bigger thing

Most of the AI tooling that's launched in the last six months has been variations on "Claude or GPT, but more so." More tools, more agents, more autonomy. Osaurus is a quieter argument: a lot of the workload doesn't need to be cloud-hosted, and the friction of running it locally was the actual blocker, not the model quality. Fix the friction and the choice changes.

I don't know if this kind of tool wins long-term. The economics for the providers point the other way; they want as much of your inference on their cloud as possible. But for designers, writers, and indie developers who care about cost or privacy or just don't want to be 100% downstream of one company, this is the first local-AI app I've installed that I'm still using after the novelty wore off.

Worth a Friday evening of your time. It's open source. The download is on the Osaurus site.

If you try it, the thing to know up front: the consumer-friendly UI is the real point. Don't go in expecting an Ollama replacement. Go in expecting the app you wished Ollama was.