Switchyard

Point Claude Code or Codex at an open-source model and the agent still expects to speak its own API. Switchyard sits in between as a Rust proxy, translating among OpenAI Chat, Anthropic Messages and OpenAI Responses formats, and the request is served by vLLM, NVIDIA NIM, Ollama or any OpenAI-compatible endpoint. The client never notices the swap.

The same proxy spreads traffic across several models for A/B benchmarking, applies signal-driven stage routing, or runs a routing algorithm you write yourself against a typed interface. Prometheus metrics cover requests, errors, latency and tokens, plus the overhead of the routing itself. It comes out of NVIDIA NeMo under Apache-2.0 and works as a library as well as a binary. The README says plainly that the project is still maturing.