Claude Code Jev Smart Router — AI skill for Claude Code

claude-code-jev-smart-router

An HTTP proxy for Claude Code that selects the Claude model per request, to reduce cost and latency.

Motivation

Claude Code serves a session from one selected model, changing it only at fixed points (see [Routing logic](#routing-logic)) rather than according to what each request requires. That model handles every request in the agentic loop, including the ones that do not need it: file reads, greps, test runs, commit messages.

Routing those requests to a lower tier has three effects:

Effect	Mechanism
Cost	Lower tiers have lower per-token prices, for input and output both.
Latency	Smaller models return sooner. Separately, lower effort reduces the number of agentic turns a step takes.
Usage limits	On a Pro, Max or Team subscription there is no per-token bill. Routing consumes less of the plan's allowance instead.

Classification adds one API call per routed request: a fraction of a cent, and well under a second of added latency. That is small enough to sit in the request path, which is what makes per-request routing possible at all. See [The classifier](#the-classifier).

Prompt cache constraint

Prompt caching limits how often switching is worthwhile. Each model has a separate cache, so a mid-conversation switch causes the new model to re-read the conversation prefix at full input price. In a long session that prefix accounts for most of the token volume, so switching on every request can cost more than using a single model throughout.

The proxy prices each switch against the cache rebuild it would cause and applies it only when it pays back within a few turns. See [Cache management](#cache-management).

Savings are unmeasured. Whether routing reduces cost on a given workload depends on that workload. See [Measuring cost impact](#measuring-cost-impact).

How it works

It listens on `ANTHROPIC_BASE_URL`, intercepts `POST /v1/messages`, extracts facts about the session from the request