Claude Code Shadow Bill — Development skill for Claude Code

claude-code-shadow-bill

Audit what your Claude Code usage would cost at API list prices, straight from the transcripts already on your disk. One file, no dependencies, read-only, nothing leaves your machine.

$ python3 whered_my_tokens_go.py --top 5

    bill  cache write   cache read  what you typed
   $7.72      767,616       15,877  "done.. single commit shows up now"
   $6.35      629,347            0  "both 1 and 2   Submit maximum"
   $6.06      601,793            0  "don't use em dashes bro as they screm AI generat"
   $5.74      568,315       15,877  "Morning, it's july 5   I tink we should do next "
   $5.18      512,749       15,877  "Good morning.. let's do clean up now? 230"

12,992 requests, 1.89B input-side tokens, 93.8% cache reads, ~$2,009 at list prices

Those are my real numbers. My most expensive API requests ever are an acknowledgment, a menu choice, and a parade of good mornings, because an agent's bill is not proportional to what you type. It is proportional to what the harness re-establishes. The full story, receipt by receipt: [Anatomy of a $5 "Good Morning"](https://karanbansal.in/blog/good-morning-tax/).

Run it

curl -O https://raw.githubusercontent.com/karanb192/claude-code-shadow-bill/main/whered_my_tokens_go.py
python3 whered_my_tokens_go.py

Python 3.7+, standard library only. It reads `~/.claude/projects/**/*.jsonl` and writes nothing. Auditing a backup or another machine's copy: `--dir `.

The three views

Leaderboard (default)

Your priciest individual API requests, with the message that triggered each. `--top 25` for more rows. Fair warning: your leaderboard will also be greetings.

`--sessions`

Which sessions ate the budget. Subagent transcripts roll up to the session that spawned them, with their share shown separately, so fan-out heavy sessions are visible at a glance:

     bill  of it subagents  requests   input tokens  session
  $379.62            $7.13     1,880    546,083,759