Auto Model Router — AI skill for Claude Code

auto-model-router

A cost-, cache- and quota-aware LLM router. It sits in front of any number of OpenAI-compatible providers and picks a model per user turn so that tasks get solved at the lowest expected cost. It can also route *whole jobs* to a coding agent's official CLI - including one running on a flat-rate subscription - so that work lands on a plan you already pay for instead of a per-token bill. What the vendors allow there is quoted, with links, in [`TERMS.md`](TERMS.md).

Status: experimental, measured. Full method and numbers: [`EXPERIMENTS.md`](EXPERIMENTS.md).

Vision

This router is one link in a longer chain, and it is built so the other links can be plugged in without changing the routing logic.

1. Evidence comes from a benchmark API. Which models and providers exist, how capable each is per topic, what a task actually costs there, and what each provider charges for a cache read or write — served by a benchmark API in the benchmarkheaven.com format, cached locally with a TTL, and every number carrying its basis and how strong that basis is (`auto_router/bench.py`). Local measurements override it, because headline scores mis-rank specific models and effort levels.
2. A Jev-class model classifies the request. Local Laya on CPU is the sample configuration's default; hosted Jev and a no-model heuristic are selectable. Topic, difficulty, whether it needs tools or a long context, whether it builds on the previous turn, and what a wrong answer would cost. Hosted Jev took about 0.6 s; the current local CPU backend is much slower (measured below). Both receive only a scrubbed and truncated summary of the turn, never the raw transcript.
3. Expected cost decides where it goes. Call cost at the route's real cache state, times the measured chance of success, plus the price of a failure.
4. **The targets are deliberately het