ctx-tax

Every skill you install charges rent in every session. This shows your bill.

Agent skills, CLAUDE.md files, rules, and session hooks all load into your agent's context before you type your first word. That is paid attention: it shapes behavior, crowds out your actual work, and nobody itemizes it. ctx-tax does.

npx @liustack/ctx-tax
# or straight from the repo, no npm needed:
npx github:liustack/ctx-tax

Every session, before your first word:

  skill descriptions (20 skills)                  2k
  ~/.claude/CLAUDE.md                           1.9k
  ~/.codex/AGENTS.md                            2.6k
  ~/.claude/rules (2 files)                       1k
  ──────────────────────────────────────────────────
  always-on total                               7.6k  = 0.8% of Claude Code 1M · 2.9% of Codex ~258k effective

When a skill fires (top by max):

  skill                               on trigger   max w/ references
  ppt-master                                 19k                301k
  vercel-react-best-practices               1.8k                 56k
  claude-api                                4.5k                 44k
  ...

Two tiers, because that is how the cost actually lands:

• Always-on: skill descriptions, memory files, and rules that every single session carries, whether they get used or not.
• When a skill fires: what one invocation loads (the skill body), and the worst case with every reference file pulled in. A skill whose max rivals the window is a skill that crowds out the actual work.

Run it from a project directory to include that project's .claude/skills, CLAUDE.md, and AGENTS.md.

What it scans

./.claude/skills, ~/.claude/skills, ~/.agents/skills, ~/.codex/skills (one copy per skill name, priority in that order), CLAUDE.md and AGENTS.md (global and project, Claude Code and Codex), .claude/rules, and it flags SessionStart hooks in settings, which inject extra unmeasured context each session.

Percentages are shown against the windows the two harnesses actually run: Claude Code's 1M, and Codex's ~258k effective (the app caps GPT-5.5 and 5.6 at a 272k catalog entry with a 95% compaction threshold, whatever the model page advertises).

The ruler

Tokens are estimated as ascii characters / 4 plus one per CJK character. Rough (about ±15%), but identical for everything measured, so the comparisons hold even where the absolute numbers wobble. Locale twins (*.zh-CN.md and friends) are excluded from a skill's max, since a session never loads both languages.

--json gives you the raw numbers.

Why we built it

We ship liustack, a four-skill working loop that bets on staying light. This tool is how we measure that bet, published so you can audit anyone's claim, including ours.

MIT.