How I Cut My LLM Costs 60% With a Local Router (Open Source)

I run AI coding agents all day. Claude Code, Codex, various tools through OpenClaw. My API bill was climbing fast. The problem: every prompt, no matter how simple, was hitting expensive models. "List the files in this directory" costs the same as "refactor this authentication system" when you're sending everything to Claude Opus. So I built NadirClaw. It's an open source LLM router that classifies prompts in ~10ms and sends them to the right model. Simple stuff goes to cheap or free models. Complex stuff goes to premium. How it works NadirClaw runs as a local proxy on port 8856. Any tool that speaks the OpenAI API can use it. You just change your base URL. pip install nadirclaw nadirclaw serve Under the hood, a sentence-embedding classifier scores prompt complexity in about 10ms. No LLM call for the classification itself. What we just shipped in v0.6.0 Fallback chains - When a model fails (rate limit, server error, timeout), NadirClaw cascades through a list of fallback models automati