I scanned my PC and found 457 AI agents — here's what I built to manage them

#cli #devtools #ai #opensource

The Discovery

I've been using AI coding agents heavily — Claude Code, Codex, Pi Agent. One evening I asked myself: how many agent-related files are actually on this machine?

I wrote a scanner. The result: 457 assets. Configs, API keys, model caches, permissions — scattered across directories I didn't even know existed.

The Problems

Discovery — No single view of all agents
Routing — Simple queries hitting expensive models
Spending — Zero cost visibility per agent
Security — API keys in plaintext, broad permissions

The Solution: AMA

I built AMA (Agent Management Application):

ama scan → find every agent
ama route → auto-select the right model
ama spend → track costs per agent/team/project
ama audit → security report

How It Works

AMA hooks into Claude Code, Codex, and Pi Agent's configuration layers. It reads their agent definitions, tool permissions, and model routing rules — then surfaces everything in a single dashboard.

It also includes a lightweight router that intercepts model calls and redirects them based on rules you define. Simple refactoring → cheap model. Complex architecture → powerful model. Security-sensitive → local-only.