Context Tax

context-tax

[](https://github.com/fomoles/context-tax/actions/workflows/ci.yml)  *English · [українська](README.uk.md)*

**Most of what you pay an AI coding agent is not thinking. It is re-reading.**

Across 6,069 real steps on one machine, **63.5% of the bill** was the model reading back context it already had. Output — the actual thinking and answers — was 18.6%.

That share is not a constant. It grows with the session, because the price of a step is a straight line in how much context the step carries. So a long session gets more expensive per step, forever, until it dies. One line in a settings file stops that.

`ctxtax` is a single Python file, no dependencies, that reads your own local Claude Code transcripts and tells you what your context is costing you — then sets the ceiling that caps it, and proves the ceiling actually fired.

---

Try it in 30 seconds

Nothing is installed, nothing is sent anywhere, nothing is modified.

curl -O https://raw.githubusercontent.com/fomoles/context-tax/main/ctxtax.py
python ctxtax.py measure --days 7

You get eight sections of your own numbers: what you spent, how the price of a step grew with session depth, the fitted price law, where the money went, what compaction costs you in wall-clock time, whether it damaged the work, and what a different ceiling would cost.

Then, if the numbers say what they said here:

python ctxtax.py cap        # writes one key, backs up settings.json first
# start a NEW session, work for a while
python ctxtax.py verify     # green only when compa