Claude Usage Audit: find where your Claude spend and time leak

Proof on our own usage

We ran it on ourselves first

Across about 5,200 prompts and 2,000 sessions of our own Claude Code work, the scanner showed this.

19%

Of prompts were one-word restarts

Things like "go", "deploy" and "try again": the agent had paused and a person had to push it on. Rules and hooks remove most of them.

97%

Cache hit rate, and usage still adds up

Even with good caching, long sessions and the wrong model for the job drive most of the usage. The report shows which.

Sessions past 500k tokens of context

Oversized sessions are the quiet cost: every turn re-reads everything. We show where to split work and what to hand to cheaper models.

What we analyze

Seven places usage leaks

✓Model mix: how much work runs on the most expensive model when a cheaper one would do.
✓Cache efficiency and context size: sessions that re-read huge context on every turn.
✓Restart nudges: the one-word prompts that mean the agent stopped when it should have finished.
✓Permission friction: rejected and repeated approvals that slow every session.
✓Tool and MCP overhead: connectors and skills that load into every turn and rarely get used.
✓Subagent use: where delegating to smaller models saves the most.
✓Bulk work that could run on a free local model.

How it works

Four steps, and your data stays with you

Run the scanner

One Python file reads your local Claude Code history and writes a summary of counts and token totals. It writes no prompt text, file paths, project names or code.

Read it, then send it

Open summary.md and summary.json yourself. If you are happy with it, send us summary.json. That is the only file we ever receive.

Get your report

A written report: where your usage goes, what is wasted, and a ranked list of fixes with the expected effect of each.

Optionally, we implement

We can apply the fixes for you: a rules file, enforcement hooks, permission settings, model and subagent routing, and local-model offload.

Try it first

The scanner is free

Run it and see your own numbers before you decide anything.

python claude_usage_audit.py --out ./audit-out

FAQ

Questions, answered

Do you see our prompts or code?

No. The scanner writes only anonymous totals. Names of custom skills, agents and connectors are replaced with short codes, and short prompts are counted but never shown unless they are a plain restart word such as "go" or "try again". You read the output before you send it.

Who is this for?

Teams and solo operators who use Claude Code or the Claude API heavily and want to know where their usage limits, spend and time go.

Does it work on any machine?

It needs Python 3 and reads the Claude Code history folder on your computer. For a team, each person runs it and we combine the summaries.

How much does it cost?

Tell us about your setup and we will send a proposal. The scanner itself is free.

See where your Claude usage goes

Run the free scanner, then tell us what you want to fix.

Request my audit →

Questions? WhatsApp us · hello@netwebmedia.com