A list of anomalies without ranking is a list of things to worry about; ranking them turns it into one thing to change. Low cache hit rate is usually the one worth reading first, because it inflates cost without showing up as unusual usage.
A local Rust server that parses the JSONL session logs under ~/.claude into SQLite and reports on them. The logs are never modified and nothing leaves the machine.
- Session logs parsed, de-duplicated and written into the local database — sync_db
- One session scored across ten dimensions — analyze_session
- Project-level aggregation with sorting, and a cross-project view — analyze_project, analyze_global
- A monthly cost report, with daily granularity available — cost_report
- Six anomaly types detected and scored by severity — HighCost, LowCacheHitRate, CostInefficient, ExcessiveToolUse, HighTokenUsage and UnusualModelMix — anomaly_scan
- Daily, weekly and monthly usage trends with burn-rate projection — trend_report
- Six skills that route a plain question to the right report, in English or Chinese
Install as a Claude Code plugin with claude plugin install claude-token-analyzer — the binary downloads itself, so no Rust toolchain is needed. Every environment variable is optional: CTA_DB_PATH, CTA_PROJECTS_DIR, CTA_ARCHIVE_DIR, CTA_PRICING_PATH and CLAUDE_CONFIG_DIR override the default paths.
One command — claude plugin install claude-token-analyzer