Cursor setup guide

1. Version and verification

Not tested yet. This setup follows the tool's own documentation.

2. Before you start

You need an account, an API key and a positive balance. Without credit a request fails with HTTP 402 (insufficient_quota).

Create a key Add credit

3. Configure

Cursor Settings → Models → API Keys

OpenAI API Key: sk-atr-YOUR_KEY
Override OpenAI Base URL: https://api.tokenwave.dev/v1
Add model: claude-sonnet-5-5

4. Choose a model

These models work with Cursor. The examples above use the one you pick here.

Browse the models that work in Cursor

5. Check the setup

Open the chat, pick the model you added and send a message. Use Ask mode for chat-completions models and Agent mode for models with a Responses endpoint.

6. Send a first prompt

Send this prompt: “Reply with the single word ready.” Then open the Requests page: the request appears within a minute, with the model, token counts and cost.

Open Requests

7. Known limits

Agent mode needs a Responses endpoint. Tab completion does not use custom keys. Cursor must reach the gateway over the public internet, which it does. Whether Cursor accepts official model IDs as custom model names is not verified yet.

8. Troubleshooting

Symptom	Cause	Fix
401 Unauthorized	The key is wrong, revoked or has stray spaces.	Copy the full key again from the Keys page and check that it is active.
403 Forbidden	The key may not use this model, or the account is restricted.	Check the key's allowed models on the Keys page, or use another key.
404 Not Found	The base URL is wrong: /v1 duplicated or missing.	Use the address from the example exactly. Claude Code takes it without /v1; the other tools with /v1.
model not found	The model ID has a typo, or your key cannot use it.	Copy the ID from the catalog, or list what your key may use with GET /v1/models.
400 Extra inputs are not permitted	The tool sent a field the chosen route does not accept, usually on a converted endpoint.	Pick a model that offers the tool's native endpoint (see its card on the model page) or update the tool.
An HTML page or “Attention Required” instead of JSON	A web firewall in front of the gateway blocked the request body.	Retry once. If it repeats, contact support with the time and the request ID.
A parameter I sent seems to have no effect	The gateway only forwards fields the route is known to support; the others are dropped, not rejected.	Look at the response header x-atr-ignored-fields: it lists the top-level fields that were not forwarded. Remove them, or contact support if you need one.

9. Undo

In Settings → Models clear the OpenAI API key and the Override OpenAI Base URL, then remove the model you added.