# auto-router

> A free, MIT-licensed LLM router. It classifies each request with Jev (TypeSafe) or a local Jev-class model, prices every model for that request from Benchmark Heaven data including warm and cold prompt-cache prices, and sends it to the route with the lowest expected cost. It runs locally with Claude Code, Codex, OpenCode, OpenClaw, Hermes Agent, GitHub Copilot or Cursor.

## Install

- [Install guide for AI agents](https://whichmodel.app.mintapis.com/agents.md): read this first; the same file is also served as https://whichmodel.app.mintapis.com/install.md
- [install.sh](https://whichmodel.app.mintapis.com/install.sh): macOS, Linux, WSL
- [install.ps1](https://whichmodel.app.mintapis.com/install.ps1): Windows PowerShell
- [Setup guide for people](https://whichmodel.app.mintapis.com/run)

## Source and evidence

- [Router repository](https://github.com/fstandhartinger/auto-model-router): code, README, TERMS.md, EXPERIMENTS.md
- [Evidence](https://whichmodel.app.mintapis.com/evidence): the paired Opus 5.5-versus-router A/B study has not run; see the replay simulation and the separate measured router run with their limitations
- [Results](https://whichmodel.app.mintapis.com/results): the measured 78-task run and the replay
- [How it works](https://whichmodel.app.mintapis.com/how)