My first Cursor test uses a tiny prompt, not a repo

A custom base URL changes where requests go

When you type a question in Cursor, the editor sends a request to an AI service. A base URL is the starting address for that service. Changing it tells Cursor to send supported requests somewhere other than the usual provider endpoint.

An OpenAI-compatible service accepts requests in a format that many AI tools already understand. That shared format can let Cursor connect to a relay without a separate integration for every model provider. It does not mean every model or every Cursor feature will behave identically.

MyRelay is a personal, pay-as-you-go relay with one key and one budget across multiple providers. Its API base is https://api.myapirelay.com. The useful starting point is not choosing the most powerful model. It is making one small request and checking what happened.

Prepare a key and a modest first budget

Create a MyRelay account, then obtain your API key and check the available balance before opening Cursor's connection settings. Start with an amount you are comfortable spending on experiments rather than funding a large coding session immediately.

Treat the key like a password. Do not paste it into a chat prompt, commit it to a project, or include it in a screenshot. If you expose it accidentally, replace it through the service's key controls.

Also check Cursor's current rules for using your own API key. A relay balance does not replace a Cursor subscription, and using a custom endpoint does not automatically make every editor feature available through that endpoint.

Find Cursor's provider and model settings

Open Cursor Settings and look for the area covering models, providers, or API keys. Names and layout can change between releases, so focus on the OpenAI API key controls and the option to override the OpenAI base URL.

Enter your MyRelay key in the relevant key field, enable the custom base URL override, and enter the endpoint required by the integration. MyRelay's API base is https://api.myapirelay.com; OpenAI-compatible clients often expect a versioned base ending in /v1. Check the current MyRelay instructions for Cursor rather than assuming whether that suffix is needed.

Avoid entering the website address or a complete address for an individual API operation. The field needs the API base Cursor will use to build its requests. Save or enable the settings, and use a verification control if your Cursor version provides one.

Choose a model by its exact API identifier

The model name in a menu is not always the identifier an API expects. Check MyRelay's current model listing and copy the exact identifier for the model you want. If Cursor supports adding a custom model, enter that identifier there and enable it for selection.

For a small initial experiment, Gemini 2.5 Flash is one option to investigate on MyRelay's menu. Check its current pricing and compatibility before choosing it. A listed model is not a guarantee that every Cursor mode supports it, especially when tools or other advanced capabilities are involved.

Test with a question that needs almost no context

Open a fresh chat, select the intended model, and ask for a brief explanation of a simple programming concept. Do not attach your repository, request broad edits, or start an agent task yet. A short exchange makes connection problems easier to isolate.

After the response, inspect the request record in MyRelay. Members can see each request's tokens and cost. Confirm that the model is the one you intended, then compare the charge with your expectations before sending more context.

Check the connection before increasing your spend

If authentication fails, recheck the key and whether the custom endpoint is enabled. If Cursor reports an unknown model, compare the identifier with MyRelay's listing. An endpoint error may indicate an incorrect base URL or version suffix. Avoid repeatedly retrying a long task while troubleshooting.

Once a small chat works, try one narrowly scoped coding question. Larger context, longer answers, and agent loops can increase spending. Keep checking request costs as you expand your usage, and remember that a working custom chat connection does not prove that autocomplete or every other Cursor feature uses the same route.