Cursor Custom API Key, Cline Endpoint and Copilot BYOK (2026)

On May 28, 2026, VS Code 1.122 made custom endpoints available without GitHub sign-in. To use your own key in Cursor, open Settings, then Models; for a gateway, enable Override OpenAI Base URL. In Cline, select Requesty or OpenAI Compatible. In VS Code, add a Custom Endpoint for Chat.

BYOK at a glance

Bring your own key (BYOK) means entering a model provider or gateway credential into your coding tool. A custom endpoint also changes where the tool sends the request. We run Requesty, the gateway used in these examples. You can also use a direct provider key.

As of October 2026, the settings and boundaries are:

Cursor's X Subscription overview says new customers start on X Subscription; legacy plans continue unless changed or canceled.

The endpoint, key and model ID belong together. A Requesty key authenticates to Requesty; an OpenAI key authenticates to OpenAI. To follow the Requesty examples, create an account first. The Model Library lists IDs in provider/model form. The Mistral and Claude models used below are paid: prepaid accounts need credits, while post-paid organizations follow their billing agreement.

Add a custom API key in Cursor

Mistral Small 4 was released in March 2026 and supports Chat Completions and function calling. Mistral's guide describes client-defined tool calls, and the Requesty deployment page lists tool calling through its OpenAI-compatible API. Cursor limits its own OpenAI BYOK description to standard, non-reasoning chat models. Test the gateway model using the checklist below.

Protocol matters for other models too. OpenAI says GPT-6.1 Terra tool calling requires Responses. Requesty's Responses reference documents openai-responses/ for native upstream Responses; its standard openai/ prefix routes upstream through Chat Completions. For Terra tool calls, use a Responses-capable client and check the exact gateway model ID.

What the override changes

Cursor says custom API keys work with chat models; Tab uses Cursor's built-in models. Compatible custom models can also be selected in the agent panel. On October 2, Cursor staff said that IDE subagents in current versions inherit the custom OpenAI key and base URL. This setup covers the IDE; see Cursor CLI configuration and Cloud Agents for those surfaces.

The override is not per model. Cursor staff explains that built-in picker models on the affected OpenAI request path can also be sent to the override. To return to Cursor subscription models, disable Override OpenAI Base URL and turn off the custom OpenAI API Key. Re-enable both for Requesty. Cursor staff also describes disabling the custom key when returning to subscription models.

For Enterprise, an admin controls Personal API Keys (BYOK) under Team Settings, then Models. Cursor's enterprise guide describes model access separately.

Check the bill and the backend hop

The provider or gateway bills the model request. On Teams and Enterprise, Cursor's Token Rate page also lists a $0.25 charge per million input, output and cached tokens for eligible third-party requests, including BYOK. Listed exempt models remain exempt with BYOK; individual plans do not pay this rate. Cursor's general BYOK page describes the charge more broadly, so check your bill or ask Cursor how it classifies a custom gateway model ID.

Cursor says requests pass through its backend for final prompt building, with the key sent to that backend for each request. Cursor's Zero Data Retention policy does not apply to BYOK. Gateway and provider data policies also apply. A localhost endpoint on your laptop is therefore not sufficient: the Cursor backend must reach the endpoint over the network. Review authentication and TLS before exposing a private service.

Set a custom endpoint in Cline

If you use different models for Plan and Act, configure and check the provider and model in each mode. Cline also documents local models, including Ollama. Its Ollama provider uses http://localhost:11434 when Ollama runs on the same machine.

With Cline Enterprise remote configuration, an admin can set OpenAI Compatible, a base URL and optional headers. Signed-in members of the VS Code and JetBrains extensions cannot select another provider once enforced; they supply endpoint credentials where required. Our 2025 Cline tutorial has a longer walkthrough; use the current settings and model IDs above.

Use GitHub Copilot BYOK in VS Code

Update VS Code and install the GitHub Copilot Chat extension. VS Code 1.122 brought Custom Endpoint to Stable and made signed-out BYOK available for chat and tools. VS Code 1.128 introduced the model options in this example. The harness controls below describe the current VS Code interface.

VS Code documents Local harness BYOK without GitHub sign-in or a Copilot plan. A hosted model still requires provider access and payment.

Local describes where the agent runs, not where the model runs. Requesty's endpoint is hosted. For BYOK in a Copilot harness session, enable the experimental chat.agentHost.byokModels.enabled setting and restart the agent host process, as described in the VS Code 1.128 release notes. Copilot CLI and cloud sessions have separate configuration; see GitHub's CLI model setup.

VS Code uses separate utility models for chat titles and commit messages. When a Copilot token is available, these default to Copilot models even if the main chat model uses BYOK. Set chat.utilityModel and chat.utilitySmallModel to an available BYOK model, or set chat.byokUtilityModelDefault to Main Agent Model. The VS Code utility settings explain precedence and signed-out behavior.

Verify the route, then test an agent

For Cursor, compare the result with its Spending dashboard.

To isolate Requesty from the editor, set REQUESTY_API_KEY in your shell and keep it out of tickets and repositories. First, list the models allowed for that key:

curl -sS https://router.requesty.ai/v1/models \
  -H "Authorization: Bearer $REQUESTY_API_KEY" \
  -w '\nHTTP %{http_code}\n'

For Cursor or Cline, test a Chat Completions request:

curl -sS https://router.requesty.ai/v1/chat/completions \
  -H "Authorization: Bearer $REQUESTY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"mistral/mistral-small-2603","messages":[{"role":"user","content":"Reply with pong."}]}' \
  -w '\nHTTP %{http_code}\n'

For the VS Code Messages path, test the Messages operation:

curl -sS https://router.requesty.ai/v1/messages \
  -H "x-api-key: $REQUESTY_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{"model":"anthropic/claude-sonnet-5-5","max_tokens":4096,"messages":[{"role":"user","content":"Reply with pong."}]}' \
  -w '\nHTTP %{http_code}\n'

Check for a text response and stop_reason. If the response stops at max_tokens, raise the limit and retry. Anthropic explains that thinking and visible text share the output limit. These commands test basic protocol requests, not streaming or the editor's tool calls. The Requesty error reference explains origin, which distinguishes router errors from upstream provider errors.

Give each developer a key

Requesty gives the team one gateway for these configured requests without distributing provider credentials to every editor. Create a separate Requesty key per developer, or per developer and tool, and set a monthly cap on each key. Requesty analytics groups usage by key, user, provider and model. You can revoke one credential without replacing the others.

Requesty logs can include message content and tool results. Turn off payload logging per key, or enable organization-wide zero data retention on Enterprise. On pay as you go, members can see organization-wide logs and analytics; Enterprise scopes visibility by role. With payload logging off, new requests retain operational metadata for the route check but have no stored content in the Messages tab. Previously captured payloads remain accessible. These controls govern Requesty's storage; provider and Cursor policies are separate.

The editor uses a Requesty key to reach the gateway. If your organization has provider contracts, it can use Requesty's provider-key settings and choose those credentials in a routing policy. Those provider keys are separate from the key entered in the editor.

For failover, create a fallback policy, choose its ordered models and use policy/your-policy-name as the model ID where the client accepts it. Each target needs to support the workflow's context, streaming and tools. Requesty attempts the next configured target on supported failures. For Claude thinking sessions, a change of model or provider account can cause earlier thinking blocks to be dropped. Check Anthropic's preserved thinking rules when designing the chain.

Requesty provides access to 600+ models. Pay as you go uses prepaid credits, with a 5% fee on top of credit top-ups and no seat fees or minimums. For example, a $100 credit top-up incurs a $5 fee: $100 + $5 = $105 charged, with $100 credited to the balance. See billing and pricing.

Which should you pick?

Pick Cline for its native Requesty provider, generic OpenAI Compatible endpoint and documented local-model path. Pick VS Code if you want BYOK chat without a Copilot plan. Pick Cursor if it is already your editor and you are comfortable keeping Tab on Cursor's model path.

Create a Requesty account and configure one editor. For VS Code, keep our GitHub Copilot guide at hand.