Connect an NVIDIA DGX Spark, or any AI box on your network

A Spark, a second PC, or any machine on your network running Ollama or another OpenAI-compatible server connects through the same "OpenAI-compatible" provider as a local one, once you give it a network address instead of localhost.

Last checked against the app on · View as Markdown

Steps

  1. On the Spark, or whichever machine will hold the model, run an inference server that speaks the OpenAI API. Ollama is the common choice, and it already answers that API at /v1 with no extra setup.
  2. That server only listens on its own machine by default, so a second computer cannot reach it yet. On Ollama, set the OLLAMA_HOST environment variable to 0.0.0.0:11434 before starting it, so it accepts connections from the network rather than only itself, then restart the server: setting the variable without restarting changes nothing.
  3. Find that machine’s address on your network, either its LAN IP or a hostname your network resolves, and note the port it is listening on.
  4. In NeuralVault, open Settings, then AI & extensions, then API keys, and find the OpenAI-compatible provider.
  5. Give it the base URL using that machine’s address in place of localhost, for example http://192.168.1.23:11434/v1.

How to tell it worked

The model picker lists the models the remote machine reports, the same way it would for a server on your own machine. A populated list means the connection reached across the network.

If it does not work

The model list is empty, but the server answers fine when you test it on the Spark itself.
Almost always the bind address. A server left on its default 127.0.0.1 refuses every connection that is not from itself, including one from NeuralVault on a different machine. Confirm OLLAMA_HOST is actually set to 0.0.0.0 and that the server was restarted after you set it.
It answers in a browser on the Spark itself but not from NeuralVault on another PC.
A firewall on the Spark is the next most common cause. Allow inbound traffic on the port your server uses, 11434 for Ollama, on whatever machine is holding the model.
You want to know if this is specific to DGX Spark.
It is not. Anything that can run an OpenAI-compatible or Ollama server on your network connects the same way: the address is what changes, never the method.