token-track

To track Claude Code token usage, you can run token-track to parse local JSONL logs. It filters duplicate streaming entries, maps usage to per-model metrics, and calculates compute units to reverse-engineer your actual usage limits.

Reverse-engineering Claude Code usage limits involves analyzing ground-truth JSONL logs to deduce the weekly compute unit cap. This tool applies Formula A to calculate burn-rate estimates and per-model cost analyses for accurate capacity planning.

Yes, you can generate burn-rate projections for Claude Code compute units. The tool reads your local JSONL history to produce per-day, per-week, and session-level breakdowns, enabling accurate tracking of your 527.6M weekly CU limit.

Yes, it works exclusively with local JSONL files to analyze per-model costs. The tool maps raw usage entries to specific models, filtering out duplicate streaming data to output structured metrics suitable for dashboards and calibration stores.

The compute unit formula calculates CU by summing per-model multipliers across input, output, cache creation, and cache read tokens. Specifically, it uses the equation CU = Σ_model[mult × (inp×1.0 + out×5.0 + cc×1.25 + cr×0.1)].

You can calibrate Claude Code model limits by processing local JSONL logs to extract structured, calibration-ready metrics. The tool supports automatic model-limit calibration and adaptive data processing for evolving prompts.