I am having an issue where the OpenAI API key toggles off on its own. I use API keys for work and my Cursor credits for personal stuff. I toggle on the OpenAI key, and within a day it has automatically toggled back off.
If I don’t manually check it every day, I’ll start eating up my personal credits without realizing it, because there’s nothing else in the UI that indicates if you’re using an API key or not. The Anthropic API key works fine and doesn’t do this.
There are other older forum posts for this bug, but they’re locked so I can’t reply.
Steps to Reproduce
Add an OpenAI API key in Cursor settings and toggle it on.
After a day or so, look in that area of Settings again and see that it has become turned off.
Expected Behavior
The API key stays toggled on until the user turns it off manually.
I’ve seen a few threads about this issue, but they were closed without a solution. Since it kept happening to me too, I decided to write a small extension to fix it myself. It basically forces the OpenAI key setting to stay “ON”.
I use “Override OpenAI Base URL” to use either Vertex or OpenRouter.
It works. however during agent execution/plan/agent mode the “OpenAI API Key” toggle gets automatically disabled without any notification and Cursor starts incurring costs and usage against my usage limits. Obviously the expectation is to honor the configuration and use the BYOK that was explicitly enabled and setup.
No notifications or warning are displayed.
This is a breach of trust!
Not only you make it hard to integrate with external BYOK providers like OpenRouter and Vertex but i feel completely sabotaged with these silent breaches of configurations that never notify users of the configuration changes. You just get slapped with the bill.
Steps to Reproduce
Setup OpenRouter or Vertex via the only possible way in the app “Override OpenAI Base URL” and get the agent to build a plan and execute. eventually the enabled “OpenAI API Key” toggle will flip to disabled without any notification and cost will start incurring.
Expected Behavior
Don’t breach user’s trust by disabling an explicitly configured toggle. At least throw an error. This is so low I cannot believe it’s a bug. It feels like a feature.