zhangrr/claude-nvidia-proxy ? reverse-engineered prompt

Build me a small Go service that lets Claude Code talk to NVIDIA models through a local proxy.

I want it to listen on port 3001 and expose a Claude style POST /v1/messages endpoint, then translate those requests into something NVIDIA accepts and forward them to the configured NVIDIA chat completions URL. Please read the API key and upstream URL from config.json, with optional environment variable overrides, and make sure I can protect the proxy itself with an optional inbound API key too.

It should support normal responses and streaming responses, and work well enough that I can point Claude at http://localhost:3001 and use models like moonshotai/kimi-k2-thinking, minimaxai/minimax-m2.7, z-ai/glm4.7, or deepseek-ai/deepseek-v3.2.

Keep it simple, use only standard Go if possible, and include clear config defaults so I can run it quickly with go run . and start using it right away.