Story · AI tooling

Metering the AI coding tools

Developer tooling · own initiative

AI coding assistants are the new power tools — and worth using deliberately. So when GitHub Copilot would not tell me how much quota I had left, I went and asked its API directly.

I work with AI coding tools daily — OpenCode, Codex, Claude — and like any tool with a bill attached, they reward being used deliberately. Copilot has a usage quota, but no useful way to watch it. You find out you are running low the way you find out a pipeline broke at hour twenty: too late.

So I put a proxy between the client and GitHub, watched the traffic, and found the undocumented API that reports quota state. On top of it I built a live-updating quota meter — a small always-visible gauge of how much Copilot I have left, which turned pacing my usage from guesswork into a glance.

Terminal quota monitor showing Copilot usage at 100% with the message: Quota exhausted! You timed that almost perfectly chief — The meter, having opinions. Landing on exactly 100% the day before reset is either luck or pacing — the tool says pacing.

It is the same move as the pipeline monitoring wall and the mainframe map: when a system will not tell you what is going on, instrument it until it does. AI tools are not exempt.

Terminal bar chart of daily Copilot quota consumption over two weeks, with green, yellow and red bars — Daily consumption over the cycle — the graph that turns “how much have I used?” into a glance.

Current experiment in the same spirit: using a small local LLM as a front-line worker that rightsizes requests — so the big, expensive models only get the work that actually needs them.

AI tooling LLMs Reverse engineering Developer experience Cost awareness

← All stories

Have a system that needs this kind of attention?

Tell us where your current testing approach breaks down. We will help you define a practical path forward.

Talk to WhileOne