One model alias. Multiple providers.

A small Go proxy for routing LLM requests, translating wire formats, and failing over when a provider cannot serve them. No control plane. No database server.

Point your client at http://127.0.0.1:8787/api/v1. Keep the same OpenRouter model names and request format; the router picks an upstream and maps the model ID.

Up and running

go build -o llm-router ./cmd/llm-router
export OPENROUTER_API_KEY=sk-or-...
./llm-router -check
./llm-router

Requires Go 1.26 or newer. Run from the repository root to use the included router.yaml. The default listener is loopback-only; configure a gateway key before exposing it to a network.

How requests flow

  1. Your client sends a request using a configured model alias.
  2. The router chooses an eligible upstream and rewrites the model ID.
  3. If that upstream fails before the response starts, the router tries another target.
  4. The response streams back in the client’s wire format.

Read why llm-router for the tradeoffs, or design notes for the routing invariants.

Edit this page