You bought the AI coding tools. Do you know what they cost you?
Not the invoice — you have that. The part underneath it: how much you pay to carry context, how often your agent re-reads files it already has, and what your always-on config charges you on every turn. cc-audit reads the session logs already on your disk and tells you.
Reads ~/.claude/projects. Prints to your terminal. Sends nothing.
Here's what it found on ours.
We ran cc-audit against 1,371 of our own Claude Code sessions — 31.7 days, $6,425 of spend. This is one developer's corpus, not an industry benchmark, and we'd rather say that than dress up a sample of one. Run it on yours and the numbers will differ. That's the point.
- 86%of the bill was carry, not output
- $5,526 of $6,425 went to re-sending context the model had already seen. Generated tokens — the part you actually asked for — were the minority of the invoice.
- 35%of file reads were re-reads
- A third of every Read hit a file already sitting in context. The counter resets on /compact and /clear, so these are genuinely redundant, not bookkeeping.
- 21,882tokens of standing context, every turn
- Config that loads before the task is even stated: 6,429 tokens of project CLAUDE.md, 1,755 global, plus skills and plugin listings. You pay it on turn one and every turn after.
- 5.4%invocation rate across 8 MCP servers
- Eight servers configured. Their tool definitions ride along constantly; they get called on roughly one turn in twenty.
- 0times 11,550 tokens of instructions were followed
- Three conditional 'read file X when Y' rules, confirmed present, weighing 11,550 tokens. Measured against what actually happened, they fired zero times.
- $123per month in spawn tax alone
- 1,032 subagent spawns, each prepended with a 15,527-token prefix. Nobody decided this; it accumulated.
None of these are a verdict on anyone's skill. They're habits with a dollar figure attached — the kind you can't see by eye and can fix in an afternoon once you can.
Habits, with dollars attached.
Not a grade. Every signal is something you can act on, measured from what actually happened rather than what you meant to do.
Context carry
The share of your bill spent re-sending context rather than generating output. Usually the largest line item, and usually invisible.
Redundant re-reads
How often the agent reads a file already in its context. Content-based, resets on /compact — so it measures real waste, not activity.
Standing context tax
What your CLAUDE.md, skills, plugins, and MCP servers cost on every single turn, before the task is even stated.
Skill ROI ledger
Per-skill carry versus what it actually earned, sorted into dead-weight, heavy-but-earning, and cheap-fine.
Context knee
The context size where your re-reads and tool errors start climbing — your personal degradation point, not a generic number.
Spawn economics
What each subagent spawn carries in prefix tokens, and what that adds up to across a month.
We'd rather you trust it than use it.
A tool that reads your dev sessions has to be boring about privacy. So: it does nothing you didn't ask for.
Runs on your machine
It reads the session transcripts Claude Code already writes to ~/.claude/projects. No account, no sign-up, no daemon.
Nothing is sent by default
There is no phone-home. The audit prints to your terminal and stops there. Three flags change that, and only if you type them: --judge and --open upload a de-identified summary, and cc-audit fix sends your config file for review.
It doesn't use your source
The analysis runs on token accounting, tool calls, and file paths — never the contents of the files you edited.
Nothing left behind
npx it once and it's gone. cc-audit fix writes proposed patches to a local folder and never applies them itself.
One machine is a curiosity. A team is a decision.
cc-audit shows one engineer their own habits, which is genuinely useful and where it ends. The harder question — whether the tools you bought are paying off across the org, and which engineers are getting leverage versus generating cleanup — needs more than one laptop's logs.
That's Promptster for Teams: the same read, across the team, without ever touching your source code. If cc-audit told you something you didn't know, that's the version worth talking about.