The release names vLLM, TensorFold and LM Studio as examples. It reads each backend's model list and routes a request to the server holding the exact requested model ID.
The router supports chat completions, completions, responses, embeddings, streaming, model-list endpoints, health and readiness endpoints, and Prometheus metrics. It preserves model names and request bodies, returns a 409 error for duplicate model IDs, and formats router errors in the OpenAI error format. It is available as signed APT packages for amd64 and arm64, a Docker image, and Linux binaries.
The release also documents optional endpoint and client API-key authentication, disabled-by-default CORS, graceful shutdown, and Grafana dashboards. It explicitly has no load balancing or failover between servers, and no built-in TLS; remote deployments require a TLS proxy. The source does not establish performance, reliability under production load, adoption, or compatibility beyond the described endpoints and server examples.