Token Damage

role: Design, implementation, testing, release
stack: TypeScript, Node.js
status: npm 0.6.0 · public repository

Problem

AI coding agents spend tokens without showing a bill. Token Damage reads the local logs of six coding CLIs and prints one receipt.

================================================
            T O K E N   D A M A G E
                 customer copy
       statement · aug 25 – sep 23, 2026
        (30 days — all claude code kept)
================================================
WORDS YOU TYPED ......................... 14,690
MODEL CALLS .............................. 7,480
TOKENS READ BY AGENTS ............ 1,178,700,000
------------------------------------------------
LIST-PRICE VALUE (API-EQUIV.) ........ ≡ $809.65
------------------------------------------------
SURCHARGE (ESTIMATE, SHOWN AS A RANGE)
  ELECTRICITY ..................... ≈ 26–120 kWh
------------------------------------------------
DAMAGE CLASS
          ┏━━━━━━━━━━━━━━━━━━━━━━━━━┓
          ┃   A C T   O F   G O D   ┃
          ┗━━━━━━━━━━━━━━━━━━━━━━━━━┛
------------------------------------------------
ADJUSTER'S NOTE
  for every word you typed, the machine read
  80,558 tokens. that's the great gatsby,
  and a third of it again. per word.
================================================
   ≡ list-price equiv · ≈ estimate · ✶ satire

customer copy · a sample month from the README, not real usage

Constraints

Local only.

It runs on your machine, with no network calls, no telemetry, and no accounts. It reads the usage logs your agents already keep, read-only; prompts are counted and never stored.

Decisions

One receipt for six agents.

One adapter per agent (Claude Code, Codex, Antigravity, Grok Build, OpenCode, and Gemini CLI for API-key users) feeds a single itemized receipt: words typed, tokens read, what the same tokens would cost at API list prices, and an electricity estimate shown as a range. The adjuster’s note, a joke from a tired insurance adjuster, is printed in red, marked as satire, and never carries a real number.

Validation

Checked against an independent tool.

Every counted line is checked against hand-derived fixture totals, per day. Before each release, real logs are cross-checked against ccusage: 83 of 83 real days matched before 0.6.0.

Outcomes

Published from CI.

Version 0.6.0 is on npm, published from GitHub Actions through npm trusted publishing with provenance. The repository is public and MIT-licensed.

Tradeoffs

A boring number beats a wrong one.

Dollar figures are list-price equivalents, not what you paid; on a subscription you pay the plan. Claude Code deletes transcripts older than 30 days by default, so a receipt covers what survived and says so.