← All articles

Features

Deep Scan: see where your Claude Code tokens really go

Deep Scan reads the Claude Code and Codex transcripts already on your machine and shows where context went — free, local, no login.

Deep Scan: see where your Claude Code tokens really go

By Ted Mavera, Product Manager at HarnessUp

Before you can cut anything from an agent session, you have to see it. That is the entire job of Deep Scan: it reads the conversations already saved on your machine and shows, item by item, where the context went. No model call, no sign-up, no estimate borrowed from someone else's account. It is the measurement layer of HarnessUp, the token-saving suite for Claude Code and Codex, and it is free.

This article explains what Deep Scan does, who it is for, how to run it, and how to read the output without fooling yourself.

What Deep Scan actually does

Claude Code and Codex already write conversations to disk as you work. Deep Scan opens those transcripts and turns them into a spend map: which rounds grew, what got resent from one round to the next, and how much of a session went to real work versus overhead. It does this without calling the model, so measuring your usage costs you nothing extra.

Three properties matter more than the charts:

  • It runs locally. HarnessUp is 100% local. Your transcripts are read on your machine, and nothing leaves it.
  • The numbers are yours. They come from your own history, calculated on your screen, and stay available for a counter-check later — including after you change a setting and want to know whether it helped.
  • It costs nothing to try. No login, no e-mail, no password. Deep Scan and 19 other measurement and monitoring tools are free, so you can calculate your waste before deciding whether a paid plan is worth it.

Deep Scan is not a one-off report. It is the top of a chain: measurement tells you which of the other tools is worth turning on, and the tools give you something new to measure.

Who it is for

HarnessUp sorts its audience into three groups, and Deep Scan is the common starting point for all of them.

  • People on a fixed plan who hit the limit. If your five-hour window runs out well before the day does, Deep Scan shows you which rounds were expensive and whether the cause was rereading, long output, or the model doing work a cheaper one could do.
  • People paying by API. When every token is on the invoice, the useful question is not "how many tokens did I use" but "which tokens bought nothing". That is exactly what a per-round view answers.
  • People who would rather not live in the terminal. The results are read in a panel, not by parsing logs. The heavy lifting happens under the harness.

If you only want one thing from HarnessUp, measurement is a defensible choice. A number you can verify is worth more than a savings claim you cannot.

How to run it

  1. Install HarnessUp. The package is npm install -g harnessup, there is a Windows download, and macOS and Linux are supported. Everything runs on your machine. HarnessUp attaches through the official Anthropic and OpenAI hooks, so it works in the CLI, the desktop app and IDE extensions, with your account or via API.
  2. Open the panel. No account required. The panel is where measurement and control live side by side.
  3. Run Deep Scan. It reads the transcripts already saved on your machine. Nothing is uploaded and no model is called to produce the report.
  4. Read your own baseline. Look at context per round with and without HarnessUp, and at the percentage of waste cut. Treat these as your starting point, not as a score.
  5. Turn something on and re-measure. In the panel, keys 1 to 8 toggle tools, and Z undoes everything. Change one thing, look at the meter, and keep the change only if the number moves.

Reading the output honestly

Two habits keep a measurement tool useful instead of decorative.

First, separate the demo from your data. The panel ships with illustrative demo values so you can see what the interface looks like before you have history. Yours come from the Deep Scan, from your transcripts, measured end to end. If a number is illustrative, the panel says so.

Second, read "up to" as a ceiling, not a forecast. HarnessUp advertises up to 78% of work tokens saved, up to 350% more token yield and up to 3x more work on the same plan. Those are the biggest gains measured in real windows, not a promise about yours. The reason the product shows you your own figure alongside is that the honest comparison is between your before and your after.

"Up to" means: the biggest gains measured in real windows. Your own number shows up in the panel, measured end to end on your own history, and stays available for a counter-check.

A quick way to read the panel:

What you seeWhat it tells youWhat to do next
Context per round, with and without HarnessUpHow much each turn carries forward, and how fast a session inflatesIf the gap is small, the waste is probably in output or model choice, not in rereading
Percentage of waste cutHow much of the avoidable spend the enabled tools removed on your historyTurn keys on one at a time and watch this move
Per-session spend and cap statusWhich sessions were expensive, and whether a cap would have stopped oneSet a Stop Loss so a runaway session ends instead of continuing

From measurement to the fix

Deep Scan is diagnostic, not therapeutic. The savings come from the tools you enable afterwards, and each one targets a different step of the bill.

  • Output. Response Diet asks the agent for a direct answer, blocks the rereading of files it already read, and cuts the giant output before it enters the context. The preamble ("Sure! I'll help you with that") is paid for once and then reread in every round that follows.
  • Reading. The reading step has its own set of keys, Delta Read among them, so the agent does not re-ingest what has not changed.
  • Compression. Smart Compress compresses without calling the model, which keeps the saving from costing you another request.
  • Model choice. Cheap Swarm decides in milliseconds which subagent does each task, with the right model and effort, across 26 task classes — search and lint on a light model, while the expensive model only decides.
  • Hard limits. Stop Loss ends the session when spend reaches its cap, so a bad loop cannot quietly consume a window.

None of this requires you to change how you prompt. There are over 200 one-click keys in the app, and any of them can be undone with one click — which is the point of measuring first. You are not asked to trust a default; you are asked to look at your own history and keep what helps.

What stays on your machine

Deep Scan reads transcripts, and transcripts contain code, paths and sometimes secrets. That makes the local guarantee the most important line in the spec. HarnessUp runs the official claude and codex with your account and states as a written commitment that it never reads, stores or passes on your token or your key. It does not intercept the API; it attaches at the official hook points, and the local layer runs with a p95 under 50 ms.

If your organization has rules about sending session data to third parties, that architecture is the answer: the analysis happens on the same disk the transcripts are already on.

Start with the number, not the pitch

Token savings are easy to claim and hard to feel. Deep Scan makes the claim falsifiable on your own machine: run it, read the rounds that inflated, enable one tool, run it again. If the number does not move, turn the tool off — the undo is one key.

That is the part I care about as a product manager. A suite that saves up to 78% of work tokens is only useful if you can see which 78% applies to you. Measurement is how you find out, and it costs nothing to look.

This article was drafted with AI assistance in HarnessUp Publish and approved by the HarnessUp team before publishing.