cursorapikey.com/Guide
Cursor API Keys Not Working: Custom Endpoint Fixes
When Cursor API keys are not working, the cause is almost always one of four things: the base URL, the model name, the key itself, or a limit on the account. A single curl command tells you which one.
Updated


Key points
- Test the key outside Cursor first with curl; if curl works, the problem is in the Cursor settings.
- Base URL must be https://api.cursorapikey.com/v1 and the model name must be exactly uncensored.
- 401 means a bad key, 404 a bad URL, 429 a rate or concurrency limit, and a model error a wrong name.
- Menu labels change between Cursor versions, so match the field purpose rather than the exact wording.
Step 1: test the key outside Cursor
Take Cursor out of the picture. If a plain request works, your key and account are fine and the fault is in the editor settings. If it fails, the error body names the problem.
curl -i https://api.cursorapikey.com/v1/chat/completions \
-H "Authorization: Bearer YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"uncensored","messages":[{"role":"user","content":"Reply with the word ready."}]}'
The -i flag prints the status line and headers. A 200 status with a choices array means success. The X-Balance-USD header shows your remaining credit, which is the first thing to check if you get an authorization or payment error.
You can also list models, which confirms the key without spending anything:
curl https://api.cursorapikey.com/v1/models -H "Authorization: Bearer YOUR_KEY" | grep uncensored
The response should contain the model ID uncensored.
Step 2: set the fields in Cursor
Open Cursor's settings and find the Models section. You need three values, and a mistake in any of them produces a failure that looks like a broken key.
- OpenAI API key: paste your key with no spaces or line breaks at either end.
- Override OpenAI base URL: turn the override on and enter
https://api.cursorapikey.com/v1. Keep the/v1and leave off/chat/completions. - Custom model: add a model named
uncensoredand enable it. Disable models you do not use so Cursor does not select one the endpoint does not know.
After saving, use the verify option if your version has one, then pick uncensored in the chat model dropdown. A fresh chat tab is a good test, since an old tab may keep the model it started with. The step-by-step walkthrough in how to use the Cursor API shows the same setup with more context.
Match the error to the fix
| What you see | Likely cause | Fix |
|---|---|---|
| 401, "invalid key", "unauthorized" | Key mistyped, truncated, or revoked | Copy it again from the key page; check for trailing spaces |
| 404, "not found" | Wrong URL, missing /v1, or a doubled path | Use exactly https://api.cursorapikey.com/v1 |
| "model not found" or similar | Model name differs from uncensored | Add the custom model with that exact name, lowercase |
429 with rate_limited | Over the requests-per-minute limit | Slow down; trial keys allow 60 per minute, full keys 300 |
429 with concurrency | Too many requests at once | Reduce parallel agent runs; trial 2, full key 8 |
429 with Retry-After: 5 | Temporary overload | Wait five seconds and retry |
| Works in curl, fails in Cursor | Setting not saved, wrong tab model, or feature that does not use custom keys | See the next section |
Failed or refused requests are not charged, so retrying while you debug costs nothing.
When curl works but Cursor still fails
This is the most confusing case, and it has a few usual explanations.
The wrong model is selected. Cursor keeps its own built-in models in the list. If one of them is selected, the request does not go to your endpoint at all. Select uncensored explicitly in the dropdown of the chat or composer you are using.
Not every feature uses your key. Features that rely on Cursor's own models, such as inline autocomplete, may keep using them regardless of the key you add. Custom keys apply where you can pick the model, which is chat and edit modes. Check Cursor's current documentation, since this changes between releases.
Concurrency from agent mode. Agent and multi-step runs can fire several requests at once. A trial key allows two at the same time. If errors appear only in agent mode, you are likely hitting the concurrency limit, and a full key raises it to eight. A streamed request holds its slot until the stream ends, so long outputs occupy slots longer.
Tool calls. The API supports tool calling, with tool_calls streamed and results passed back as role: "tool". Quality of tool use depends on the model, so if an agent run stalls on a tool step, try the same task in plain chat first.
Check the account side
If curl returns an error about balance, the key is valid but the credit is empty or the trial has ended. The trial is $0.50, valid for 7 days, claimed with one click on the API key page after sign-up. After that you need a top-up.
Top-ups are any whole amount from $20 to $500, with +5% from $50, +10% from $100 and +15% from $200. You can pay by crypto (USDT on TRC20, USDC on Base). The pricing page shows the table. If you have no key yet, get one here.
Keep one key per tool. If you pasted a key into several editors and scripts, rate limits are shared across them, which can look like random failures.
Script check for a stubborn case
If you still cannot find the cause, run a short script that mirrors what the editor does. It uses the OpenAI SDK, which Cursor-style clients are modeled on, and it prints the exact error object.
from openai import OpenAI, APIStatusError
client = OpenAI(base_url="https://api.cursorapikey.com/v1", api_key="YOUR_KEY")
try:
r = client.chat.completions.create(
model="uncensored",
messages=[{"role": "user", "content": "Write a one-line Python hello world."}],
stream=True,
)
for chunk in r:
print(chunk.choices[0].delta.content or "", end="")
except APIStatusError as e:
print(e.status_code, e.response.headers.get("X-Request-Id"))
print(e.response.text)
If the script streams text, the endpoint, key and model are correct, and the remaining issue is a Cursor setting. If it prints a status code, use the table above. Keep the X-Request-Id value when you contact support, since it identifies the exact request. Full parameter details are on the documentation page.
Keep it from breaking again
A few habits prevent most repeat problems. Store the key in one password manager entry and copy from there, rather than from chat logs or notes where line breaks sneak in. Write down the three values (base URL, key, model name) so you can re-enter them after an editor update, since updates sometimes reset overrides. After each update, send one short chat message before starting real work.
Watch the balance. The X-Balance-USD header is returned on each response, and you can also read it from GET /v1/balance. Credit never expires, so a modest top-up sitting in the account is not wasted. If you use the key from scripts as well as from the editor, give long batch jobs a retry loop that waits five seconds on a 429 and stops after a few attempts, so a script cannot starve your editor of its concurrency slots.
Cost check and what to read next
A coding chat request with 10,000 tokens of file context and a 700-token answer costs 10,000 / 1,000,000 x $0.15 = $0.0015 for input and 700 / 1,000,000 x $0.50 = $0.00035 for output, about $0.00185 in total. Rates are $0.15 per 1M input and $0.50 per 1M output tokens. The article on Cursor API pricing covers monthly estimates.
If the request that failed was a large one, remember the window: prompt plus completion must fit in 100,000 tokens, and the request body is capped at 8 MB. Attaching a whole repository to a chat can exceed both. Trim the context to the files that matter and try again.
More guides
Questions and answers
Why does Cursor say my API key is invalid?
Usually the key has a stray space or was copied partially, or the base URL still points at another service. Re-copy the key, then confirm the override URL is https://api.cursorapikey.com/v1.
Which model name do I enter?
Enter uncensored, in lowercase. It is the only model ID the API offers.
Does autocomplete use my key?
Not necessarily. Some Cursor features use Cursor's own models, and custom keys mainly apply to chat and edit modes where you choose the model.
Are failed requests billed?
No. Failed or refused requests are not charged, so retrying while you debug costs nothing.
Your key is one form away
Create an account, copy the key, change the base URL. That is the whole setup.