Guide

How to see your Claude Code usage at a glance

Last updated 31 August 2026

Claude Code has a session window and a weekly allowance, and both run out at inconvenient moments. Knowing where you are without stopping to ask changes how the day goes.

What actually runs out

There are two limits, and they behave differently:

• The session window. A rolling block that starts with your first message and resets a fixed number of hours later. Heavy use empties it quickly.
• The weekly allowance. Slower, and usually the one that surprises people — often split so that the largest model has its own share separate from everything else.

They reset independently. Being fine on one says nothing about the other, which is why a single number is not enough.

Checking it inside Claude Code

The CLI reports the state of the current session and where it is against the weekly figures. That is exact and always available, and the drawback is the obvious one: you have to stop and ask, and you tend to ask only after something has already gone wrong.

Putting it somewhere you can see

What is worth surfacing continuously is small:

• How much of the session window is used, and when it resets
• How much of the week is used, across all models and for the largest one
• What today has cost
• Which projects are running, and which have finished and are waiting on you

That last one matters more than it sounds. If you run several sessions at once, the expensive mistake is not a limit — it is a finished task sitting unnoticed for an hour.

Reading it from your own account

The figures belong to your account, so anything showing them has to be signed in as you. In practice that is an OAuth sign-in, the same as any other app connecting to a service — you approve in a browser once, and the tokens are refreshed after that without asking again.

Two things to know if you build this yourself. A static API key will not read plan usage; it is a different kind of credential from the one the session endpoints expect. And tokens expire, so anything that fetches once and caches forever will quietly stop updating after a day or so — usually without saying anything, which is the worst way for it to fail.

What a good glance looks like

Percentages, not raw token counts: nobody has an instinct for what 4.2 million tokens means, and everybody has one for 61%. Show the reset time next to the number, because "61%" and "61%, resets in twenty minutes" are different situations. And keep the running-projects list short — the point is to notice a change, not to read a log.