Cursor does not see your models after changing the base URL

Cursor connects to a third-party API through its OpenAI key settings. The setup is simple, but it has a few non-obvious places where nearly everyone stumbles.

The URL needs /v1

Cursor speaks through the OpenAI SDK, which treats the versioned prefix as the base. The correct value is:

Without /v1 requests land at the root and you get a 404. This is cause number one — check it first.

You must press the verify button

Cursor will not use a third-party key until verification succeeds. Filling the fields and closing settings is not enough: the toggle stays off, the IDE keeps using its own backend, and you believe you are connected.

The order is: paste the key, paste the URL, press verify, wait for the green result, and only then close the dialog.

Model names must be added by hand

Cursor does not fetch the model catalogue automatically. The list keeps its built-in names, which may not exist on a third-party gateway — a request goes out for gpt-4o, we have no such model, and back comes a 404 about a missing model.

Add the names you need manually, exactly as they appear in the catalogue. To see the exact identifiers:

curl -s https://api.zukko.pro/v1/models

It is worth removing the leftover built-in names: picking one by accident produces a confusing error.

Not every Cursor mode uses your key

This is a limitation of the editor, not the gateway. Ordinary chat and in-editor edits go through the API you configured. Some proprietary features — inline autocomplete, certain agent modes — run on Cursor's own infrastructure and never touch your key.

Hence the familiar complaint: "my usage history is empty although I worked all day". What is billed is what actually passed through the gateway, not everything you did.

Confirming it worked

Ask anything in chat, then open the usage history in your dashboard. An event means the connection works. No event but an answer means Cursor replied from its own backend — go back to the verify-button step.