1stStep.ai

Open-source MIT skill

Suggest a lighter model tier. Confirm before it runs.

Auto Model Router is a portable policy for Cursor, Claude, Codex, and other multi-model agents. It classifies work into the lightest sufficient tier, shows the suggestion, and waits for you to confirm or override. The skill stays usable without an email—this page is only for optional feedback.

GitHub repository INSTALL.md MIT license

Example flow (illustrative)

How the skill chooses a tier

A model conversation that shows classify → suggest → confirm → run. Not a live IDE recording or Cursor embed.

First example: rename getUserName to fetchUserName. Auto suggests fast because it is a clear rename with low judgment. Confirm, then it runs on the fast tier. Second example: auth refresh fails only for users who signed up before token rotation. Auto suggests reasoning. Confirm, then it runs on the reasoning tier.

  1. Rename getUserName → fetchUserName across the auth module. Auto suggests fast — clear rename, low judgment. Confirm, then it runs on the fast tier.
  2. Auth refresh fails only for pre-rotation users. Auto suggests reasoning — unknown-root-cause debugging. Confirm, then it runs on the reasoning tier.
Classifyfast, standard, reasoning, or max — capability labels, not vendor model names.
Suggest → confirmThe agent explains the tier first. You confirm or override before substantial work.
Escalate onceIf a light attempt fails, it says so and steps up. Explicit model choices always win.

Honest scope

This is a readable heuristic, not trained routing ML and not a billing dashboard. Usage savings happen only if the confirmed lighter tier is mapped to a cheaper or faster model (or lower effort) and that mapped choice is what actually runs. Percentages in the repo estimator are relative comparisons from a local log, not vendor invoices or promised savings.

Leave feedback

Email is optional. The skill works the same either way. If you share an email, you must opt in to hear about this tool and related 1stStep tools.

Leave feedback, an email, or both. At least one is required.

No silent email harvesting. Required only if you want a reply or updates. See the privacy policy.