Cursor Local Model: Connect a Self-Hosted LLM

Cursor can use a local model for chat through its “Override OpenAI Base URL” setting. Cursor sends those requests from its own servers, so a localhost address is unreachable: the model needs a public HTTPS URL, and prompts and code context pass through Cursor on the way. This tutorial connects a Cursor local model served by Marigold, states which Cursor features use it, and covers VS Code with Continue as the route where requests go straight from the editor to your machine.

This assumes the setup guide is done and the chat example is running.

How Cursor reaches a custom model

• Requests to a custom base URL come from Cursor’s servers, not the editor. Tools such as curxy exist for this reason: they tunnel a local server to a public URL.
• Tab completion always uses Cursor’s built-in models. Agent and Edit rely on Cursor’s own models; tool calls may fail against a self-hosted model even when chat works (ngrok: connect a coding agent to custom models).
• With the override on, all OpenAI-family requests go to your endpoint, including models picked from Cursor’s built-in list (Cursor forum, staff reply, 21 August 2026).

Cursor changes this behaviour between releases. Check it against the version you run.

1. Check the endpoint locally

curl http://localhost:8000/v1/models

Copy a model id from the response, then send one chat request:

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "<model id>", "messages": [{"role": "user", "content": "Say hello."}]}'

2. Give it a public URL

Marigold requires no API key and accepts requests from any caller. A public tunnel therefore exposes the model to anyone who has the URL.

For a short test, a Cloudflare quick tunnel prints a temporary trycloudflare.com address:

cloudflared tunnel --url http://localhost:8000

Stop the tunnel when the test ends. For regular use, put an authenticating proxy in front of the API that rejects requests without the key you set in Cursor.

3. Configure Cursor

In Cursor Settings, under Models:

1. Turn on OpenAI API Key and enter a value. Marigold ignores it; an authenticating proxy checks it.
2. Turn on Override OpenAI Base URL and enter https://<tunnel host>/v1.
3. Select Add model and enter the model id from step 1.

Select that model in a chat and send a message.

What to expect

Two Marigold limits apply to chat. stream=true returns correctly framed chunks without token-level streaming, so each reply appears at once. Only one tool call per assistant turn round-trips correctly.

To return to Cursor’s own models, turn the base URL override off.

Keeping code on your machine: VS Code with Continue

Continue, an extension for VS Code, calls the configured apiBase directly, so a local LLM in VS Code needs no tunnel and no third party sees the prompts. Add the model to Continue’s config.yaml:

name: Local config
version: 0.0.1
schema: v1

models:
  - name: Marigold
    provider: openai
    model: <model id>
    apiBase: http://localhost:8000/v1
    apiKey: unused

The same OpenAI-compatible endpoints serve any local AI IDE or client that accepts a custom base URL. Clients that need token streaming or parallel tool calls will hit the two limits above.

Changing the model

The model comes from the running package’s models.yaml. Adding a new model covers the edit, rebuild and populate loop.