LLM Model Router plugin
The model router (llm-model-router) gives a route one base URL for many providers. The client names a model as <provider>/<model>, for example openai/gpt-4o, or by an alias with no /. The plugin picks the route’s connections that serve that model and speak the endpoint the client called, tries them in order until one answers, and relays the answer. It rewrites only the model field of the request body. It doesn’t translate between provider formats.
Where it runs
The failure mode is fixed: a request that the plugin can’t handle is never forwarded.
Configuration
Groups the route’s connections under a provider name. The client uses that name as the model prefix. You choose the name, and it doesn’t have to match the connection’s provider.
fallback tries the connections in list order. loadbalance picks the first connection by weight, and then tries the rest in list order.
The connections in the group. Each has a name and, for loadbalance, a positive weight. A fallback group takes no weight.
Maps a model name with no / to an ordered list of targets. Each target has a provider group and the model ID sent to it. Later targets are fallbacks.
A connection serves exactly the models on its allowed models list, so each connection in a group needs that list. The console has no form for this plugin, so you enter the configuration as JSON. Fabric checks it when you save an enabled instance. A disabled instance isn’t checked until you enable it. For the checks, see Fix a configuration that Fabric refuses.

