vbc1406/flux-router ? reverse-engineered prompt

Build me a self hosted AI proxy that sits in front of my model calls and automatically sends each request to the cheapest model that still meets the quality I need.

I want to be able to point an existing OpenAI compatible client at it with almost no code changes, and have it handle routing, retries, and fallbacks behind the scenes. It should also track spend across a whole agent run, so I can set a max budget and max number of steps, and have the system switch to cheaper choices as the run gets close to the limit instead of blowing past it. Please include a simple local dashboard to see usage, routing decisions, and cost history, plus a way to run it locally or in Docker. If you need to check current provider docs or API details, look them up online.