Connect Cursor

An existing Cursor installation can use our public API through Override OpenAI Base URL. This applies to supported chat requests. Cursor Tab and other Cursor services keep their own requirements. See Cursor’s API-key help.

Have your service key and an exact model ID from our models endpoint ready. This setup uses Chat Completions. Cursor currently documents OpenAI BYOK support for standard, non-reasoning chat models; a model appearing in our catalog does not establish Cursor support.

1. Save the connection in Cursor

Open Cursor Settings → Models and find API Keys. Keep a record of any connection you intend to replace.

Save the key and URL. Use the API base, without /chat/completions appended. Enter the key in Cursor’s settings, never in a chat message.

Cursor sends these requests through its servers. The configured endpoint must be reachable over public HTTPS; localhost and private LAN URLs do not work directly. Our public Base URL is the relevant value for this setup. Cursor staff describe this requirement.

2. Add and select the model

In Models, enter the exact service model ID in the model search field and use Add Custom Model, then select it in the chat model picker. Do not add a local provider prefix or invent an alias: Cursor sends the model ID to the endpoint.

Refresh Model List refreshes Cursor’s catalog; it does not discover our /v1/models. Add service models manually. If an ID already belongs to a built-in OpenAI model, select that entry with the override enabled instead of creating a duplicate. Cursor’s setup guidance and model-ID explanation cover these behaviors.

3. Verify a new conversation

Start a new chat, select the intended model explicitly and ask:

Reply with CONNECTION_OK. Do not use tools or edit files.

Check the reply and the matching request in your service dashboard. A saved key or successful settings check does not prove that a chat used our API. The test consumes API balance.

Limits and troubleshooting

• The OpenAI URL override does not redirect every built-in vendor model. In particular, built-in Anthropic models follow their separate provider route; adding their name does not make this setup work. Use a compatible model that Cursor actually routes through the override. Cursor staff explain the routing limit.
• An unavailable model error requires checking the exact ID and key permissions. Do not rename the requested ID to bypass a name collision.
• Agent work additionally needs compatible tool calls and streaming; a text reply alone does not verify them. Reasoning and vendor-specific parameters may still be incompatible.
• To return to your previous provider, switch off the custom OpenAI-key route and restore your previous model selection. Recheck the saved URL when enabling the override again.

Cursor’s API-key help also explains that the key passes through Cursor’s backend for request processing. For service-side failures, use our API troubleshooting reference.