/route — Token-Saving Model Router

Route each task to the cheapest model that can actually do it, instead of burning frontier tokens on work a small/local/subscription model handles fine. Two mechanisms:

• Delegate (works live, no restart): keep Claude as the brain, shell out a sub-task to a cheaper CLI or local model, bring the result back.
• Base-switch (whole session, needs restart): repoint Claude Code's base model to a cheaper provider for an entire session — same mechanism as /free-mode.

When to use

• /route or "route this", "do this on the cheapest model", "run it locally", "save tokens / save $", "which model for X?"
• Before a batch / mechanical job (bulk refactors, find-replace-at-scale, boilerplate, test generation)
• Before a huge-context read (summarize a giant file / repo) that doesn't need deep reasoning
• When the work is private and shouldn't leave the machine

Don't route a single genuinely-hard reasoning task — just run it on the frontier model. Routing has overhead; only pay it when the task is cheap-able.

Step 1 — Detect what's available

Run this first and read the result (caches a profile to ~/.claude/skills/route/providers.json):