Select your ~/.claude folder to load data

Your browser does not support persistent folder access. Use the folder picker below.

Guide

This guide explains how to use the Claude AI Usage Report, where the data comes from, and how the analysis works. All processing happens locally in your browser — your data never leaves your device.

Features overview

Dashboard — Developer and Manager mode overviews with budget, cost trends, ROI, and quality metrics.
Detailed Usage — Daily and weekly usage tables with limit highlights.
Tools & Skills — MCP tool and skill usage, plus tool call frequency.
Data Summary — Detailed session list and impact analysis from usage facets.
Settings — Configure your dev rate, daily limit, working days, and time-estimate baselines.

How calculations work

Important: what is estimated vs. what is measured

The time saved, lines changed, dev value saved, and ROI figures are rough heuristic estimates — not measurements. There is no recording of actual time spent, no comparison against a baseline without Claude, and no empirical study backing the numbers. Treat them as indicative only.

What IS reliable

What IS reliable: token counts, API costs, session counts, success rates, helpfulness scores, and outcome distributions — these are read directly from your local data files.

Time & lines saved (per session) — estimated

Each session's primary_success type is mapped to a baseline (e.g. multi_file_changes → 45 min, single_file_fix → 15 min). These values are arbitrary and not derived from measurement. They are then multiplied by a complexity factor min(3, 1 + goalCount × 0.15) based on the number of goal categories — also heuristic. Only fully_achieved / mostly_achieved sessions count toward totals.

Dev time value saved — estimated

Estimated minutes saved ÷ 60 × your effective hourly rate. Since the time saved figure is itself an estimate, this monetary value compounds that uncertainty. The rate is either your hourly input or derived from monthly salary as monthly ÷ (hrs/wk × 52 ÷ 12).

% Time saved — estimated

Estimated hours saved ÷ (contracted hrs/wk × weeks of data) × 100. The denominator is measured from the facets file date range and is reliable; the numerator is the heuristic estimate described above.

Token cost — measured

Read directly from the JSONL session transcripts under ~/.claude/projects/ and matched to each facets session by ID. Pricing per model (per 1M tokens): Sonnet $3 in / $15 out · Opus $5 in / $25 out · Haiku $1 in / $5 out.

ROI — partially estimated

Dev time value saved ÷ total token cost. The cost side is accurate; the value side is the heuristic estimate above. The ROI figure should be read as an optimistic upper bound, not a precise measurement.

How time & value estimates are calculated

Time Saved (heuristic estimate)

Baseline for session type (e.g., 15 min for single_file_fix) × complexity multiplier (1 + goals × 0.15, capped at 3×). More goals = harder session = more time saved. Example: 15 min × 1.75 = ~26 min for a fix with 5 goals. Only successful sessions count.

Dev Value Saved (estimated)

Time saved (converted to hours) × your effective hourly rate. Shows the monetary value of what Claude helped you accomplish. Set your rate in Settings to match your cost or market rate.

Token Cost (actual from Claude logs)

Real API cost extracted from your Claude JSONL session logs (based on input/output tokens and model used). When '—', the session's token data wasn't captured in logs. This is what you actually paid to Claude.

Outcome

Fully achieved or mostly achieved sessions count toward totals above. Other outcomes (partially, no value, etc.) are shown but don't contribute to aggregate time/cost savings.

Data sources

The ~/.claude folder contains JSONL session transcripts under projects/ and usage facets under usage-data/facets/.

Optionally load .claude.json for skills and MCP tool metadata.

Everything is parsed in your browser. No data is uploaded to any server.