Skip to content

Share one AI server across your team

Instead of every workstation running its own model, you can run one ullama server on a capable machine (typically one with a good GPU) and have each person's Understand connect to it. The team shares a single model, a single warm cache, and one set of hardware requirements.

This uses the bundled server. If you'd rather share LM Studio or Ollama, see Set up a more robust local AI server — the client-side steps are the same idea.

The host machine sets the ceiling for everyone

Whatever model you run on the shared host is what the whole team experiences — a small model on a modest machine will feel quite different from a large model on real server-class hardware, both in response speed and answer quality. Since one choice now affects every client instead of just one workstation, it's worth picking deliberately — see Choose a model for the tradeoffs before you settle on one.

1. Run the server on the network

On the host machine, start ullama bound to all interfaces instead of just 127.0.0.1, pointing it at whichever GGUF model you've decided the whole team will share (see the note above) — for example the larger Qwen3-8B model, on a host with the GPU memory to run it well:

ullama -m /path/to/Qwen3-8B-Q4_K_M.gguf --host 0.0.0.0 --port 56767 -c 64000 --parallel 2

--parallel sets how many requests the host serves at once, and -c is the total context divided among those slots — so the example gives each of two slots 32000 tokens. Raise --parallel for a larger team, and raise -c to match, as far as the host's memory allows.

For a machine that should always be available, run it as a service so it starts at boot and restarts on failure — see Run the AI server as a background service.

Make sure the host's firewall allows inbound TCP on the port you chose (56767 by default).

If the server fails to start because its model won't load, see The AI model file failed to load.

2. Point each client at it

In each user's Understand:

  1. Open Tools → Options (Windows/Linux) or Understand → Preferences (macOS) → the Data page.
  2. Click the ... button next to the AI Provider row and choose Understand Remote.
  3. Enter the server's host and port — e.g. ai-box.example.com:56767 (the port defaults to 56767 if you leave it off).
  4. Tick the data acknowledgement and click OK.

With Understand Remote selected, Understand connects to your server and does not start a local one, so all the inference happens on the shared host. Command-line batch generation can use the shared server too — pass -server host:port to und ai.

Trusted networks only

The bundled server has no TLS, and no authentication unless you set one: --api-key <key> makes it require that key on every request. Anyone who can reach the port can otherwise send prompts to it. Only bind it to the network on a network you control, behind your firewall — never expose it to the public internet. Prompts include code context, so treat the link like any other internal service carrying source. For the full privacy picture see What can the AI access?.

If the shared server becomes unreachable

There's no client-side "test connection" button — you'll find out it's down when a request fails or AI features stop responding. When the connection drops, Understand retries automatically every 5 seconds for as long as AI stays enabled in settings; you don't need to reopen the dialog or restart Understand once the host comes back.

There's no automatic fallback to a local model. If the shared host is down, unreachable, or the network path is blocked, every client configured for Understand Remote just keeps retrying the same host — it will not silently start its own local server instead. If you need AI to keep working during an outage, switch the affected machines back to Understand Local temporarily (same Data page), or run the server as a background service so outages are rarer in the first place.