Claude Usage Monitor

Your Claude Code usage over the last — days, generated locally from ~/.claude/projects. The rolling 5-hour and 7-day windows track how close each session puts you to your cap.

generated: — range: — turns: —

—

show big number as
calibrate from /usage
%

5-hour window · anchored

—
— spent
— of cap used
—
—

7-day window · anchored

—
— spent
— of cap used
—
—

Rolling 5-hour window

—USD
— tokens · — turns
— cap: not set · set from observed peak

Rolling 7-day window

—USD
— tokens · — turns · since —
— cap: not set · set from observed peak

Throttle markers

When you get rate-limited, click the corresponding point in the scatter below. Once marked, the minimum cost across your throttle events replaces the observed-peak estimate as the gauge cap — so future readings show % of a cap you've actually hit, not a guess.
mark right now clear all

Per-turn context

Every assistant turn, plotted by time and context size. Y is the context visible to the model (cache_read + cache_write + input). Sidechain turns (subagents) are drawn smaller and dimmer so they don't dominate the eye.

models: |
scroll or drag the bottom bar to zoom · click any point to mark a throttle event

Hourly rollup

Turns per hour (bars, right axis) and rolling context average (lines). The thin line is the raw hourly average; the thick line is a 24-hour rolling average that flattens session churn into a trend. Headless and sidechain turns are excluded.

Where did the cost come from? Per-day aggregates from your entire ~/.claude/projects archive, not just the 21-day dashboard window. Use the metric switcher to spot inflection points — a sustained climb in tokens per turn usually means sessions are accumulating more context; a drop in cache hit rate means turns are paying full re-cache cost; a rising opus share moves the cost-per-token multiplier 5× upward.

metric:
Generated by claude_usage.py — data never leaves your machine · tool page