Connect LM Studio, or any other OpenAI-compatible server

Anything that speaks the OpenAI API on a URL you control connects through the "OpenAI-compatible" provider, and needs a base URL rather than a key.

Last checked against the app on · View as Markdown

Steps

  1. Start your server and note the address it prints. LM Studio’s local server defaults to http://localhost:1234/v1; llama.cpp and vLLM print theirs on launch.
  2. Load a model in that server first. An OpenAI-compatible endpoint with no model loaded answers requests with an error, which reads in NeuralVault like a connection problem rather than an empty server.
  3. In NeuralVault, go to Settings, then Intelligence, then API keys. The provider is called "Local gateway" in the model picker, and its row under "Providers that run locally" says which address it is reached at. The built-in address is http://localhost:1234/v1, which is LM Studio’s default, and it needs no key.
  4. If your server is at another address or port, or on another computer, the app has no box for it in this build. Settings shows the address but cannot change it. Put it in the file .neuralvault-compat.json in your Windows user folder: type %USERPROFILE% in File Explorer’s address bar, create or open that file, and make it hold {"base":"http://192.168.1.60:1234/v1","model":"your-model-id"}, using your own address, including the /v1 if your server uses one, and a model id your server lists. Choosing a model in the picker afterwards keeps the address.
  5. NeuralVault runs on Windows only. The server can run on a Windows PC, a Mac or a Linux machine. For a server on another computer it must accept connections from the network and not only from itself, which is set in that server’s own options, not in NeuralVault.

How to tell it worked

The model picker lists the models your server reports. If the list is populated, NeuralVault reached it.

If it does not work

The model list is empty but the server is running.
Nearly always a missing or extra /v1 on the base URL. Open <your base URL>/models in a browser: if that returns JSON, the URL is right. NeuralVault reads the same address.
It is on another computer and nothing is listed.
Check the address in .neuralvault-compat.json is that computer’s network address and not localhost, and that the server there is set to listen on the network. A firewall between the two is a common cause.
It works in a browser but not from NeuralVault.
Some servers bind to 127.0.0.1 only and some to all interfaces. Use exactly the host the server printed, not a substitute you assume is equivalent.