03 / SELECTED WORK — Agent tooling

agent-router.

A shell-based router for Claude Code and Codex. It classifies new CLI tasks without a separate model call, keeps each host's configuration and state separate, and can fall back when a rate limit is reached.

THE PROBLEM

What it tackles

Choosing a model for every new coding task is repetitive, and rate limits can interrupt the workflow.

THE APPROACH

How it works

Classify prompted tasks into size tiers without spending a model call.
Keep Claude Code and Codex installation, configuration, and state separate.
Report observed model changes and support rate-limit fallback.

WHY I BUILT THIS

“Take the guesswork out of model selection while making the choice easy to understand.”

Explore source Watch real CLI demo Try the illustration

RECORDED WALKTHROUGH / AGENT-ROUTER

Watch a task find its route.

Real agent-router dry runs classify sample prompts for Claude and Codex, then show how a tight budget changes the Codex choice.

Recorded from the real agent-router CLI with example configuration and isolated budget state. Dry runs make no model calls; model names and availability are examples. Playback is a recording, not an interactive terminal. On a narrow screen, swipe inside the player to see the full terminal.

What happens in this recording

Classify a quick lookup for Claude with ccr -n.
Classify a system-design task for Codex with cxr -n.
Set the example budget state to CRITICAL and repeat the Codex dry run.
Compare the effective tier and model chain without launching either AI CLI.