AGENTS.md token audit cost: the audit is free, the calibration is the work.

Nothing. The analyzer at ccmd.dev runs entirely in your browser, takes about 250 milliseconds on a normal-sized file, and never uploads the file anywhere. No signup, no email gate, no rate limit on the first scan. The paid tier ($9 to $19 per month solo, $49 per team) is for continuous monitoring (weekly drift email, PR diff comments, per-engineer cost attribution), not for the audit itself. The audit itself is and will stay free.

The audit prints estimatedCostPerLongRunSession, which is totalTokens times 30 turns times $15 per million tokens, divided by one million. A 1,600-token file prints $0.72. A 3,200-token file prints $1.44. A 6,400-token file prints $2.88. The math is at src/lib/analyzer.ts lines 263-269. That number is the upper bound for one specific path: Codex CLI on uncached turns, 30 turns long, Opus 4.7 at the base input rate. Your actual bill is that number multiplied by your loader's fire rate.

Because the cost path at src/lib/analyzer.ts line 264 sets estimatedTokensFireEveryTurn equal to totalTokens, unconditionally. The detector at line 41 returns 'agents.md' or 'claude.md' but the cost code never reads that value. The audit is doing the worst-case math: the whole file fires every turn at the highest input rate of the three Anthropic models. Calibrating that number to your loader is a manual step you do after reading the report. There is no toggle in the UI because Cursor versus Claude Code versus Codex versus 'CLAUDE.md with @AGENTS.md' is a four-way matrix and the audit prints one number.

Because the rubric and per-line findings are loader-independent. Cursor will not bill you a single token for AGENTS.md if you have not wired it up, but the file is still part of the repo, still gets reviewed, still drifts, and still gets copied into a new project. Auditing it for clarity and for the Karpathy rubric is useful even when the per-session dollar number is zero. The dollar number is only one of the audit outputs; the other outputs (rubric pass count, vague-line flags, bloat warnings, missing failure-mode coverage) are what most teams actually act on.

Cut the file in half. The audit cost is linear in totalTokens. Run the audit, sort the findings by tokenSavings descending, and remove the top five lines. Most hand-written AGENTS.md files have a 200-300 token block of motherhood statements ('always write clean code', 'always think before changing') and one or two 80-character workflow paragraphs that should be three bullets. Removing those typically drops the file by 40-50%. The audit will reprint a smaller dollar number on the next scan.

It is a midpoint. A short bug-fix session is 5-10 turns; a long refactor session is 50-100. The audit uses 30 because it is roughly where the empirical median falls in the Claude Code billing data shared in public threads. For a long-running session, double the printed number. For a short one, halve it. The constant is exposed at src/lib/analyzer.ts line 266 (TURNS = 30) and is the easiest knob to tune mentally; the dollar number scales linearly.

Indirectly. The audit flags lines as 'vague' (R-class), 'aspirational', 'bloat', or 'missing why', and Claude has the same problem with those lines as a human reader does: they are too soft to act on. So a line like 'always write clean code' gets flagged as aspirational with a tokenSavings estimate and a suggestion. Whether Claude in fact ignored that line on a given turn is a different observation, made by reading transcripts; the audit cannot read transcripts. But the lines it flags are the lines most likely to be ignored.

Paste each file separately. The browser-side analyzer does not chase @-imports because it cannot read your filesystem. If your AGENTS.md does @docs/style.md and @docs/db.md, the audit prints the cost of AGENTS.md plus the inlined import bodies only if you paste them inline. Otherwise the printed number is the cost of the wrapper file alone, which underestimates the real per-turn cost. The fix is to paste the resolved file (run a quick script that inlines the imports) and audit that. For Claude Code specifically, Anthropic resolves @-imports up to a hop limit before injecting into the system prompt; you want to audit the resolved blob, not the wrapper.