Can Cursor use a local model on a Mac? Not directly, and here is why

I make Outlier, a local AI app for Mac. Until today our developer page listed Cursor among the tools you could point at Outlier's local API. For Cursor that is wrong, and I have changed the page. This is what Cursor does with a model address of your own, what you can do about it, and what to use if the point is to keep the model and your code on your Mac.

Why a localhost address fails in Cursor

Most coding tools that take a model address run on your Mac and call that address themselves. Cursor does not. When you add your own API key, or change the base URL, the request goes to Cursor's servers first. Cursor's help page for API keys says "all requests are routed through Cursor's servers for final prompt building." The server then calls the address you gave it.

That is why 127.0.0.1 and localhost fail. Those names mean this computer, to whichever computer reads them. When Cursor's server reads one, this computer is Cursor's, and nothing of yours is listening there. A reply on Cursor's forum from February 2026 puts it the same way: localhost is not supported directly, and local-network addresses will not work because the server cannot reach them. A symptom that fits is a local server whose log stays empty, because nothing ever connects to it.

What your own key covers in Cursor, and what it does not

The same help page lists limits that apply even when the address does work. Custom API keys only work with chat models. Tab completion continues using Cursor's built-in models. Cursor's Zero Data Retention policy does not apply to requests that use your own key, so the data terms are the ones of whoever is behind the key. The page does not mention local models at all.

So even a perfect setup would give you your own model in Cursor's chat panel and nothing for Tab completion.

The tunnel workaround, and why I would not use it for a private model

The forum reply suggests exposing the local server as a public https address with ngrok or Cloudflare Tunnel, then pasting that address into Cursor under Settings, Models, Override OpenAI Base URL. I did not try it. I would not, if I were running a model locally to keep my code private.

The model would be local, but the path would not be. Your prompt goes from your editor to Cursor's servers, out to the tunnel's public address, and down to your Mac. The address is public too. A plain Ollama server has no password, so anyone who learns the address could use your Mac to run models.

Outlier would refuse a tunnel anyway. Its engine answers only requests addressed to this Mac: a Host header of 127.0.0.1, localhost or [::1]. Anything else gets a 421 reply. A tunnel normally passes its own public name as the Host, and rewriting it would switch off the check that stops a web page you visit from using your model. I did not try a tunnel with Outlier. I did test the check itself: a request with Host: example.com got the 421. The evidence is on what 127.0.0.1:8766 is.

What to use instead on a Mac

Pick a tool that runs on your Mac and makes the call itself. Cline is an open-source coding agent for VS Code. Zed is an editor that lets you point its AI features at an LLM API address. Aider runs in the terminal. Each takes an OpenAI-style base URL and a key. A request from your Mac to a model on your Mac never leaves it.

A local model will not match the cloud models Cursor uses by default, and predictive completion is the gap you will feel first. The Cursor alternatives page goes through that gap and the options.

Point those tools at Outlier

1. In Outlier, open Settings, go to the Advanced tab, find Developer, and switch on OpenAI-compatible API. The app asks you to confirm.
2. Copy the endpoint, http://127.0.0.1:8766/v1, and the API key. The key is hidden until you press Show, and Copy puts it on the clipboard.
3. In your tool, add an OpenAI-compatible provider with that base URL and key. For the model name use outlier-compact for the Core tier, or outlier-nano for the fastest one.
4. Leave Outlier open. The engine stops when the app quits.

The API is part of Pro, $249 once. On the free plan the toggle opens an upgrade card instead, and a request to /v1 comes back 402. The endpoint answers from the model alone: no web search, no memory, no project files. The developers page has working code and says which request fields are honoured, and how to run a local OpenAI API on a Mac has the longer version.

What I did not test

• Cursor itself. Nothing here came from running it. Cursor changes often and the forum reply is from February 2026. If Cursor has added a way to call local models, tell me at matt@outlier.host and I will fix this page.
• A tunnel, with Outlier or with anything else.
• Cline, Zed or Aider against Outlier's API. The endpoint uses the OpenAI request shape and the developers page has working OpenAI SDK code, but I did not set up these three for this page.
• Any Outlier version before 1.11.918 for the Settings path.

Frequently asked questions