What does it cost an AI agent to read your repository — and how much of that is waste?

Free, open source (MIT), runs entirely locally. No tokenizer dependency.

$ uvx contextcost .
# no install, no config, results in seconds

Measured on 17 well-known open-source repositories

Repository	Tokens to read	After proposal	Saved	Share
moby/buildkit	14,600,569	1,529,637	13,070,932	89.5%
jesseduffield/lazygit	5,766,190	1,278,993	4,487,197	77.8%
sharkdp/bat	53,715,389	23,737,443	29,977,946	55.8%
astral-sh/uv	8,855,618	4,331,842	4,523,776	51.1%
astral-sh/ruff	20,978,634	10,531,462	10,447,172	49.8%
plotly.js	63,831,059	37,008,917	26,822,142	42.0%
dask	4,315,000	2,308,363	2,006,637	46.5%
pandas	10,105,577	7,929,282	2,176,295	21.5%
rclone	7,889,081	6,169,673	1,719,408	21.8%
trufflesecurity/trufflehog	4,456,181	3,011,760	1,444,421	32.4%
gitleaks	300,980	199,238	101,742	33.8%
keycloak	18,687,556	17,290,337	1,397,219	7.5%
pydata/xarray	2,133,276	2,007,117	126,159	5.9%
restic	1,054,989	1,005,182	49,807	4.7%
astropy	7,881,727	7,669,606	212,121	2.7%
mikefarah/yq	420,446	417,457	2,989	0.7%
contextcost itself	175,189	91,872	83,317	47.6%

The "after" number is not arithmetic on guesses: contextcost proposes cuts, then walks the repository again with the proposal applied. Estimates carry a measured error bound (see ERROR_BOUND in estimate.py — currently ±23%, measured over real source, config and lockfiles); --accurate gives exact cl100k_base counts (on plotly.js the estimate landed 0.7% off).

The spread is the finding

moby/buildkit — 89.5%, 13.1M of 14.6M tokens: a vendored vendor/ tree plus generated protocol code. An agent fixing a Dockerfile frontend issue could lose 89% of its reading budget and miss nothing.
dask — 46.5%: one lockfile (pixi.lock, 932K tokens) is 22% of the entire repository's context cost.
mikefarah/yq — 0.7%: a disciplined repository correctly gets told it is clean. The tool does not invent waste.

Make it permanent

$ contextcost . --write-ignore
# accept the proposal, confirm the saving once

$ contextcost . --fail-over 8000000
# then let CI fail when the budget creeps back

Or gate pull requests automatically with the GitHub Action — every PR gets a comment with its token delta. There is also an MCP server mode so Claude Code, Cursor and Codex can call it directly.