tokstat 1.0.2

Aggregate and analyze Claude Code token consumption from local JSONL transcripts

claude-token-usage

CLI tool that aggregates and displays Claude Code token consumption. Scans ~/.claude/projects/ JSONL transcripts, estimates costs using live LiteLLM pricing data, and prints everything in a color-coded terminal table.

Installation

pip install tokstat

Requires Python 3.10+. No dependencies.

Usage

claude-token-usage                        # overview for today
claude-token-usage --period all           # all time
claude-token-usage --period "7 days"

Modes

Default — aggregated overview

Displays consumption by period, by project, by model, output speed, and grand total.

claude-token-usage
claude-token-usage --period all

--prompts — per-exchange detail

Per-exchange breakdown: user text, model, turns, tokens (input/output/cache), tool calls, cost.

claude-token-usage --prompts
claude-token-usage -p --period "7 days"

--audit — behavioral anti-pattern detection

Scans assistant transcripts for 11 categories of behavioral anti-patterns (French + English). Summary tables show breakdown by category and by model with incident rates.

claude-token-usage --audit
claude-token-usage -a --period "30 days"

Abbr.	Category	What it detects
Gaslt	Gaslighting contextuel	Denying previous statements, rewriting history
Anthr	Anthropomorphisme / fausse empathie	False emotions, fake experience claims
Hedge	Dilution par prudence	Dense hedging clusters in a single sentence
Lazy	Paresse intellectuelle	Deflecting to docs, generic non-answers, filler
Overc	Aplomb trompeur	Confident assertions followed by tool errors
Sycop	Flagornerie / sycophancy	Excessive praise, performative agreement
Compl	Acquiescement performatif	"You're right, but..." patterns
Prem.	Solution prematuree	Declaring victory before verification
Loop	Boucle d'echec	User reports same failure 3+ times
Verb.	Verbosite creuse	Long structured response to short question
FakeU	Comprehension feinte	"I understand" without addressing the issue

--anomalies — technical anomaly detection

Detects unusual patterns in per-exchange token data. Results grouped by project.

claude-token-usage --anomalies
claude-token-usage --anomalies --period "30 days"

Anomaly	Trigger	Severity
Runaway cost	Prompt costs 10x+ the P90	HIGH
High cost	Prompt costs 5x+ the P90	MEDIUM
Tool storm	30+ tool calls in a single prompt	HIGH >60, MEDIUM >30
Turn spiral	API turns 5x+ the P90	HIGH >10x, MEDIUM >5x
Cache thrashing	High cache writes with <50% read-back	MEDIUM
Context bloat	Input/output ratio >50:1 with >10K input	LOW
Empty exchange	5+ turns but <100 output tokens	MEDIUM

Thresholds are computed dynamically from your own data (median, P90).

--plan — plan & optimization recommendations

Cost breakdown by model, plan recommendation, and data-driven optimization advice.

claude-token-usage --plan
claude-token-usage --plan --period all

  All time — 17 active days / 30

  Model              Calls     Cost   Avg/day  Projected/mo  Cache  Share
  ─────────────────  ─────  ───────  ────────  ────────────  ─────  ─────
  claude-opus-4-6      321  $475.19  $15.84/d    $475.19/mo    96%   100%
  claude-sonnet-4-6      8   $0.811  $0.027/d     $0.811/mo    96%     0%
  TOTAL                329  $476.00  $15.87/d    $476.00/mo    96%

  Plan (based on All time)
    Max 20x ($200/mo) strongly recommended.
    Projected API cost: $476.00/mo — you'd save ~$276.00/mo

  Optimization Recommendations

  Model selection
    100% of spend is on claude-opus-4-6. claude-sonnet-4-6 is 2x cheaper.
      - Use claude-sonnet-4-6 for simple tasks
      - Switching 30% would save ~$49.78/mo

--export — conversation export

Exports all exchanges to a JSON file.

claude-token-usage --export
claude-token-usage --export out.json --period "7 days"

{
  "tool": "Claude Code",
  "model": "claude-opus-4-6",
  "timestamp": "2026-04-08T...",
  "user": "the user prompt text",
  "assistant": ["response 1", "response 2"],
  "turns": 25,
  "tools_used": {"Bash": 3, "Read": 7, "Edit": 2},
  "tool_errors": ["error message"]
}

Filters

All modes support --period:

--period <period>    all, hour, "5 hours", today, yesterday, "7 days", "30 days", year
                     default: today — partial match works ("7" = "Last 7 days")

Pricing

Model pricing is fetched from LiteLLM's model pricing database and cached locally at ~/.cache/token-usage/litellm_prices.json for 24 hours. Falls back to stale cache if fetch fails.