spendwatch

One project ate 51% of my AI coding spend.
No dashboard could show me.

Anthropic will never show you your Cursor usage. Cursor will never show you Codex. spendwatch reads the logs your agents already write locally and answers the question no vendor can: where is the money actually going?

npx spendwatch
$ spendwatch

  spendwatch  ·  127 sessions  ·  24,824 turns  ·  Feb 5 → Aug 27, 2026

  SPEND
    API-equivalent cost               $4,408.82

  BY PROJECT
    ~/work/atlas-api               $2,259.76   51.3%  ██████······
    ~/work/beacon                   $1,032.95   23.4%  ███·········
    ~/work/portal                     $867.11   19.7%  ██··········

  cache reuse                           54.9×  healthy

Rules it refuses to break

100% local

No network code exists in the package. Your prompts and code are never read, stored, or sent. ~700 lines — audit it in one sitting.

Never guesses a price

Unknown models are counted and reported unpriced, never estimated. gpt-5.3-codex is not gpt-5, despite the string prefix.

Never double-counts

Session replay inflated our own test number 2× — deduplication on request id is built in.

Reads

Claude Code transcripts · Codex rollouts · Cursor's commit-authorship DB (read-only). Zero runtime dependencies. Node ≥ 22.5. MIT.