How to Connect Cursor, Cline & Roo Code to a Third-Party LLM API Relay (2026)

2026-09-29 ｜ RouteAPI Team ｜ 8 min read

Cursor, Cline, and Roo Code are the three most widely used AI coding tools today, and all three bill at official rates by default: Cursor through a monthly subscription, Cline and Roo Code by burning the official API quota on your own account. Point their requests at an API relay instead, and the same models are billed at 49% of the official reference price — with one key covering 30 models across Claude, GPT, and Grok.

This guide configures each tool step by step, then walks through the problems people actually hit.

1. Why Route IDE Tools Through a Relay

• Cost: official APIs bill per token, and heavy use easily runs past $100 a month. A relay bills at 49% of the official reference price, so the same workload costs roughly half.
• Model access: some models are hard to get approved directly depending on region or account type. A relay offers one entry point, and a single base_url reaches every model it carries.
• One key for everything: no juggling three accounts, three bills, and three rate limits across OpenAI, Anthropic, and xAI. Your editor, CLI, and scripts share a single key.
• Pay as you go: top up and spend it down. No monthly minimum, and no subscription quota ceiling.

Every configuration below uses these three placeholder values — swap in your own:

base_url: https://api.route-api.site/v1
api_key:  sk-your-routeapi-key
model:    claude-sonnet-4-6

2. The Three Tools Compared

3. Configuring Cursor

Go to Settings → Models (the shortcut is Cmd + Shift + J on macOS and Ctrl + Shift + J on Windows / Linux).

1. On the Models page, find the OpenAI API Key field and paste sk-your-routeapi-key;
2. Turn on Override OpenAI Base URL and enter https://api.route-api.site/v1;
3. Click Verify to check connectivity (Cursor sends one validation request upstream);
4. On the same page, use Add model to register a model name manually, for example claude-sonnet-4-6, then select it in the model dropdown;
5. Head back to the editor and send a message — a normal reply means you're connected.

4. Configuring Cline

Cline is a VS Code extension that sends every request straight to the provider you configure, so a relay can take over completely — it is the least troublesome of the three.

1. Install Cline from the VS Code Marketplace and open the panel from the Cline icon in the sidebar;
2. Click Settings (the gear icon) at the top right of the panel;
3. Set API Provider to OpenAI Compatible;
4. Set Base URL to https://api.route-api.site/v1 — the /v1 matters;
5. Set API Key to sk-your-routeapi-key;
6. Set Model ID to claude-sonnet-4-6. It must match the catalog name exactly;
7. Tick Enable streaming and leave it on;
8. Save, then describe a task in the chat box. Cline presents a Plan first and moves into Act once you approve it.

To use Anthropic's native protocol for the Claude family instead, switch API Provider to Anthropic and set the Base URL to https://api.route-api.site — without /v1 this time. Everything else stays the same.

5. Configuring Roo Code

Roo Code is a fork of Cline, so the settings map almost one to one. The difference is its multi-mode architecture.

1. Install the Roo Code extension and open the sidebar panel;
2. Click Settings → Providers;
3. Set API Provider to OpenAI Compatible;
4. Set Base URL to https://api.route-api.site/v1, API Key to sk-your-routeapi-key, and Model to claude-sonnet-4-6;
5. Turn on streaming and save;
6. Assign models per mode: Architect handles design work and is worth the flagship claude-opus-5; Code handles implementation, where claude-sonnet-4-6 is the best value; small edits can drop to claude-haiku-4-5.

Roo Code keeps one configuration per mode, so switching modes switches both model and prompt. That makes it easy to reserve expensive models for the parts that genuinely need reasoning, like architecture discussions.

6. Common Pitfalls

1. The model name must match exactly: capitalization, hyphens, and date suffixes all count, and a typo returns model not found or a 404. Always take names from the model catalog — claude-sonnet-4-6 and claude-haiku-4-5 cannot be abbreviated from memory.
2. Anthropic and OpenAI protocols use different endpoints: the OpenAI-compatible protocol uses https://api.route-api.site/v1 (the client appends /chat/completions), while the native Anthropic protocol uses https://api.route-api.site (the client appends /v1/messages). Pick the wrong protocol, or add or drop /v1, and you get a 404 or 401.
3. Streaming has to be on: with streaming disabled in Cline or Roo Code, long responses tend to time out at the proxy layer and get cut off mid-answer. Cursor is best left on its default streaming behaviour too.
4. CORS and proxies: VS Code extensions like Cline and Roo Code use Node's network stack and are not bound by the browser same-origin policy. Move the same request into a web app, however, and CORS blocks it — you need to forward from a server. Corporate HTTPS-intercepting proxies and self-signed certificates cause handshake failures for the same reason; whitelist the domain or disable certificate interception.
5. INSUFFICIENT_BALANCE is not a key problem: an error containing INSUFFICIENT_BALANCE (usually with HTTP 402) means the balance is empty, not that the key is invalid — do not keep regenerating keys. Top up on the recharge page; the balance is credited automatically after payment, then retry the same request. Only a 401 points at the key or base_url.
6. Watch the trailing slash: do not add a slash to the end of the Base URL. Some clients concatenate /v1/ with the following path into a double slash and return a 404.

7. Why RouteAPI

• 30 models: the main Claude, GPT, and Grok families sit in one model catalog, all reachable with a single key — including coding favourites like claude-sonnet-4-6;
• 49% of the official reference price: every model is billed at 49% of the official reference price, and the discount basis is printed on each model page so you can check it line by line;
• Per-token metering: no monthly fee and no plan tiers. The console's usage records show the input and output tokens and the exact charge for every call;
• OpenAI + Anthropic compatible: Cursor uses the OpenAI protocol, Cline and Roo Code can use either, and the same key works across all of them;
• Automatic balance crediting: pay online on the recharge page and the balance is credited automatically after successful payment (usually within a minute) — no manual review.

8. FAQ

Are the relay's models the same as the official ones? Yes. You are calling the same upstream models with identical protocols and output, and the model names match too — claude-sonnet-4-6 refers to the same model on both sides.

Can one key serve Cursor, Cline, and Roo Code at the same time? Yes, and usage from all three rolls up into the same account. If you want separate bookkeeping, create additional keys in the console and assign one per tool — billing works exactly the same way.

It is configured but just spins or never responds. How do I debug it? Check four things in order: whether the Base URL carries the right /v1, whether the model name matches the catalog exactly, whether the API key was copied in full, and whether streaming is enabled. Then confirm the balance is sufficient.

How is billing calculated, and can I see the details? Billing is per token, and every model is priced at 49% of the official reference price. After each call, the console's usage records show the input and output tokens along with the amount charged.