Cursor and Continue.dev against your own DGX Spark endpoint
This page gets Cursor's chat and Continue.dev's IDE panel answering from a model on your own DGX Spark instead of their default cloud backends. You set a base URL, a key and a model name in each editor's settings. For terminal and agent tools (Cline, aider), see the coding agents guide.
You need:
- A running GPUwerk instance with an OpenAI-compatible endpoint: a console model preset such as Qwen3 Coder 30B, or your own engine on a base image.
- For a console preset, an API key with the
inferencescope, created in the console. - Cursor or Continue.dev installed.
- No SSH key. GPUwerk configures the machine, and the console's browser terminal works without one. Add your own key only if you want your own SSH client, scp or port forwarding.
Start with a running endpoint
Both editors need an OpenAI-compatible /v1 endpoint reachable from your laptop.
- Pick the endpoint. With a console model preset, copy the HTTPS endpoint from the instance page (it looks like
https://app-<instance-id>.gpuwerk.com) and add/v1; requests need your inference-scope API key. With your own vLLM server on a base image,<spark-host>in the samples is your instance's hostname,<name>.gpuwerk.com. - Protect your own engine. The instance has its own public IPv4 address, assigned by GPUwerk, so a server on
0.0.0.0:8000is open to the internet. Start vLLM with--api-key, or bind it to127.0.0.1and use an SSH tunnel, which needs your own SSH key. - Confirm it answers before you touch either editor:
curl http://<spark-host>:8000/v1/models
For a preset, callhttps://app-<instance-id>.gpuwerk.com/v1/modelswith-H "Authorization: Bearer $GPUWERK_API_KEY". You should see JSON listing the model ID, for exampleqwen3-coder-30b.
Cursor
- Open Settings, Models in Cursor.
- Add a custom OpenAI API endpoint and fill in the base URL, key and model. For a preset, use its HTTPS endpoint plus
/v1, your inference-scope key andqwen3-coder-30b:
Base URL: http://<spark-host>:8000/v1 API Key: not-needed-unless-you-put-a-gateway-in-front Model: Qwen/Qwen3.8-27B-FP8
- Select the custom model in a chat and send a test prompt. You should get an answer from your instance.
A custom endpoint usually covers Cursor's chat and edit requests. Its tab completion uses a separate model slot that you may not be able to redirect, and this changes between releases, so check Cursor's current model settings. The model name must match /v1/models exactly; a mismatch fails with an error that looks like an auth problem.
Continue.dev
Continue uses the same config.yaml entry in its VS Code and JetBrains sidebars as in the coding agents guide. Follow those steps, with apiBase pointing at your Spark or your LiteLLM gateway. Sidebar chat, inline edit and autocomplete then all use the model that entry names.
Other IDE assistants
If an assistant's settings have an "API base URL" or "custom provider" field next to the API key, the same pattern works: point it at your Spark's /v1 URL, or at a LiteLLM gateway for per-user keys and usage tracking. Support and menu locations change between releases, so check the tool's current documentation.
JetBrains AI Assistant and several VS Code extensions follow the same shape. Check the scope case by case: some plugins send only their chat panel to a custom endpoint and keep other features, such as commit message generation, on their own built-in model.
Rolling this out across a team
One developer can use the Spark's URL directly. Once several people share the box, put LiteLLM in front. Each developer gets their own virtual key, you can rate-limit a runaway agent loop before it slows everyone else down, and revoking one person's access leaves everyone else's settings alone. In each editor, the base URL changes to the gateway's address and the model name becomes the alias you set in LiteLLM's config.yaml.
What actually breaks
- Code still reaches the vendor's cloud. Cause: inline autocomplete runs on a separate system from chat, and the custom endpoint covers only chat. Fix: check what the setting covers before you assume code stays on your instance.
- An auth or network error on every request. Cause: usually a model name that does not match the server. Fix: copy the ID from
/v1/models. On a preset, also check the key has theinferencescope. - Long prompts fail or get cut off. Cause: open files, recent edits and the system prompt fill the context window before your question. Fix: raise
--max-model-lenon your own engine; the Qwen3 Coder 30B preset is fixed at 32,768 tokens. - A setting is not where this page says. Cause: both editors ship often and move menus. Fix: check the tool's current documentation.