Deep Scan
Basic FREEScans the conversations already saved on the machine and shows where context was spent, one by one. You start the read whenever you want. It doesn't change the transcripts.
Every token leaves a trail.
Your limit ran out. The explanation doesn't have to be a mystery. See where the context went, conversation by conversation, with numbers from your own history.
No login. No email. Your transcripts stay on your machine.
the receipt you were missing
in your conversation ↙
> following the token trail_
refactorcache
CSS tweakoutput
new functioninput
illustration · local reading · no files changed
your agent's receipt
Context has a cost.Now you see where it comes from.14tools to understand spending
12included in Basic FREE
0sign-ups to start
100%local, on your computer
Open the conversation.
Find the weight.
A total alone hides the story. Pick a sample session and separate input, cache and output. What uses the most tokens doesn't always cost the most.
Illustrative history / 3 sessions
click or use keys 1, 2 and 3 ↗
API refactor
18 turns · sample model
- New input
- US$ 0.72
- Cache read
- US$ 4.32
- Cache written
- US$ 0.40
- Output
- US$ 1.80
Here, cache read weighs most on the bill. The start of the conversation went back to the model over several rounds.
Demo with fictional data. This site doesn't access your history. In the app, measurements come from local sessions; the cost shown depends on the model and usage mode. US$ values do not represent a fixed-plan charge.
12lenses on your usage.
lenses on your usage.
All free.
[ open receipt ]
History becomes evidence.
First, find out what happened. No economy module switched on.
Harness History
Basic FREEThe receipt for each closed conversation: total cost and how the spend breaks down. Find the sessions that weighed most without adding it up by hand or waiting for the invoice.
Inside the Wallet
Basic FREESplits by model: new input, cache read, cache written and output. You see which band weighs on the bill, based on what your own sessions recorded.
[ invisible weight ]
Before your “hi”, context is already there.
Harness, MCPs, skills and cache also go into the conversation. See the size of that baggage.
Cache Cycle
Basic FREEShows how much reading the cache cost and when that cost started. A block placed at the start can be re-read in later rounds; the screen helps you understand why a long conversation got expensive.
Openness Meter
Basic FREEOpens the context of the first turn: what went in, where it came from and how much each built-in MCP server weighs. Find the installed baggage before you even ask the first question.
Harness Meter
Basic FREEMeasures the core Claude Code loads before any request and compares its versions. When an update increases that weight, the difference becomes visible.
[ fair comparison ]
Choose with your own history.
The right model and the right rule depend on the work you do.
Insights
Basic FREERecalculates real sessions under another rule: a cut, a cap or a context diet. Compare the simulated scenario with the recorded spend before deciding to adopt the change.
Model Insights
Basic FREE24 questions about cost, speed, errors and reasoning of the models in your use. Each answer shows the math, so you can check where it comes from and compare who works best on your day to day.
HarnessUp Bench
Basic FREEEvaluates models on Assertiveness, Optimization and Spend, with a verdict per request. See who gets it right more often and what each right answer costs. Contributing to the public benchmark is optional and anonymous.
[ well-used plan ]
The limit becomes learning.
Understand your requests, the wasted windows and the credit that slipped by.
Work Analytics
Basic FREEGathers what you ask the agent, what works and what HarnessUp does in its place. Use this picture to recognize requests that work and improve the next ones.
Bad Use
Basic FREEPoints out 5-hour windows spent with no return and explains why, from a simple task on an expensive model to a subagent that used too much. Each case comes with guidance to avoid repeating it.
Plan Usage
Basic FREEShows how much each week and each window yielded until the limit renews. See whether credit expired unused and adjust your pace to get more out of the plan.
+ 2 commands in Ultra
Take the analysis to the chat.
The diagnosis and the simulation can also start inside the conversation with the agent.
See Ultra/diagnostic Ultra
Ask for an x-ray of past usage, right in the conversation. The command points out the problems and where spend concentrated, without leaving the agent to look for the diagnosis.
/simulate Ultra
Test a plan, model or habit change on real history. See the hypothetical scenario with your numbers before making the change stick.
The math is on display.
The starting point is what your sessions record. Measurement, simulation and evaluation need to appear under different names.
Cost per band =
tokens recorded × applicable rate
÷ 1,000,000When the rate is per million tokens.
Session total = sum of the bands.
tokens recorded × applicable rate
÷ 1,000,000When the rate is per million tokens.
Session total = sum of the bands.
Start at the source.
Deep Scan reads the history available on the computer, on demand. The analysis starts from local transcripts, without modifying the conversations. What isn't recorded shouldn't become a measured number.
Separate the bands.
New input, cache read, cache written and output have different costs. The conversion to US$ uses the rate for the model and usage mode. Token count and cost need to be read together.
Read US$ in the right context.
On the API, the calculated cost can be compared with the matching charge. On a fixed-price subscription, that value is not an extra bill. Plan Usage and Bad Use help you understand how well the windows were used.
Don't confuse scenario with result.
Insights and /simulate recalculate hypotheses about the history. The simulated difference is not money already saved. Model scores and verdicts are assessments, not raw token counts.
Compare the same base.
To evaluate a change, check sessions, model, period and comparable criteria. A result observed in your history describes that slice; it doesn't guarantee the same gain on every task.
Before the first scan.
Have a question about your case?
help@harnessup.dev
Do I have to pay to see my numbers?
The 12 tools in the metering panel are in Basic FREE. You install and measure with no login. The /diagnostic and /simulate commands are part of Ultra. Basic FREE covers 1 Claude Code account and 1 project.
Does the scan change my conversations?
No. Deep Scan reads the history already saved on the machine, on demand, and does not modify your transcripts. On Basic FREE, the layer measures; the paid savings modules are not applied.
Do I need to send my data to measure?
No. Measurement runs locally. HarnessUp runs the official claude and codex with your account and never reads, stores or passes on your token or key. Contributing to the public benchmark is optional and anonymous.
Is US$ on screen an extra charge?
No. On a fixed subscription, the calculated value helps you understand how heavy your use is, without changing your harness fee. API users can compare the calculation with the matching charge. The usage mode makes a difference in how you read it.
Is the demo reading my computer?
No. The three examples on this page are fictional and exist to explore the interface. To analyze your own transcripts, install HarnessUp and run Deep Scan in the local panel.
measure first. decide later.
Your history already has the answer.
Install, open Deep Scan and see your usage before choosing a paid plan.
One line to start with npm.
Prefer clicking? The Windows installer is on the install page.
npm install -g harnessup