Fails with an unpaid OpenAI key or with no key at all

[geoffjay]

This might just be a limitation of Cursor, but it seems like this does not work without a paid OpenAI API key. If I try using the curxy URL without a key I get an error when I update the settings, and if I use a key that was created on an unpaid account I get the error The model xxx does not work with your current plan or API key. If that's just how it is and an API key on a paid account is what's required it would be helpful to call that out in the README.

Should the URL from curxy work without an OpenAI API key? If so how do you get it to use the URL without the key?

[njuFerret]

same issue, not working with "The model xxx does not work with your current plan or API key"

[ryoppippi]

@geoffjay @njuFerret
Thank you for let me know.
I investigated and found that you guys are right!!

So please send us a PR on README!

[jincdream]

#27 (comment)

[grav]

I think the issue is a combination of two things:

• Cursor does not enable overriding the URL before it has verified the key at least once
• The verification request against Ollama results in a HTTP 403 on Ollama (see log below)

$ ollama serve
...
[GIN] 2025/01/31 - 10:47:44 | 403 |       3.791µs | 2a05:f6c7:5650:0:dc2c:3a34:5f9b:d58a | POST     "/v1/chat/completions"

I proxied the request that Cursor makes against the overridden URL with Charles, and copied it as Curl. I then reduced the headers until I got a HTTP 200 from ollama.

Seems the culprit is this header: -H "origin: vscode-file://vscode-app" If that one is removed, it works.

Workaround:

1. ensure that the OpenAI models are enabled
2. Remove any URL override in the Cursor settings, and add a valid OpenAI key.
3. Verify the key
4. Start curxy with the same key, eg OPENAI_API_KEY=sk-***** deno run -A jsr:@ryoppippi/curxy
5. Override the url in Cursor settings to the one curxy returns
6. Enable only the model(s) that ollama supports

[alexellis]

Hi @grav I found your comment here by trying to do something very similar.

I paid some credits to OpenAI, verified the key, then added my override for the URL to an inlets tunnel.

When I click Verify, I see a success message in the logs of ollama:

[GIN] 2025/02/19 - 21:12:42 | 204 |      28.343µs |       127.0.0.1 | OPTIONS  "/v1/chat/completions"

However, the error showing a curl statement is still shown in Cursor.

[ryoppippi]

@alexellis @grav Hi sorry for my response.
If you don't want to pay a dollar for OpenAI, you could make a map of gpt models and local llm models. cursor uses gpt-3.5 and 4o-mini for trivial tasks, so it is assumed that you need the OpenAI key at that time.
What do you think of this feature? It would be a good idea to create some kind of configuration file so that openAI models can be bypassed by ollama.

[AmosDinh]

You can actually set the open api key such that it has only certain permissions, i believe you can restrict it so it cant call a model.