01
What is my usage costing?
Put recorded tokens into context with a cost estimate. See the rate source, and keep estimates separate from your provider’s bill.
Recorded usage. Clearly labelled estimates.Usage Radar.
Less guesswork.
More building.
Your AI coding tools, in one clear picture. Understand tokens, costs, and cache activity. Keep your prompts and code private.
Free for everyone. No account or card.
Built around the tools
you already use.
- Claude Code
- Codex
- Gemini CLI
The product, in focus
One workflow.
Three perspectives.
Recorded activity, model costs, and context reuse. One place to make sense of your own workflow.
See the whole week.
Start with recorded activity. See your daily estimates without confusing them with your provider's bill.
Know your model mix.
Compare the models you use. Understand where recorded activity and estimated costs come from.
Make reuse visible.
See cached input alongside new input and output. Understand the counters without exposing your conversations.
Workspace overview
Illustrative sampleEstimated usage cost
$128.80USD · sample data
15.0M tokens recorded3 models
A daily view of recorded activity, with estimates clearly labelled.
Your own numbers appear after you connect. Estimates are not invoices.
A useful daily habit
Less digging.
More understanding.
You don’t need another wall of numbers. Start with the three questions that help you understand your own workflow.
02
Which models do I use most?
See how your activity breaks down across agents and models. Understand what changed when you try a different tool or workflow.
One place for your model mix.03
How much context is being reused?
See cache reads and writes alongside input and output tokens. Understand reuse without sharing the prompts behind it.
Cache accounting, without the guesswork.When a session fights back
See what failed.
Say what was wrong.
Usage Radar counts the requests that failed or were retried — rate limits, overloads, dropped connections — from the transcripts already on your computer, and states them against the requests it counted, because twelve failures reads differently on a 40-request day than on a 400-request one. It cannot tell you whether an answer was correct: it reads token counts, never the answer. That part is yours to flag, and your flags are drawn beside the failures instead of being mixed into them.
Reliability · one weekSample data · not your usage
Reliability · sample week83requests that failed or were retried · 6 of 7 days with agent activity43 failed · 40 retriedof 1,657 requests · 5.0 per 100
Your quality log4answers you marked wrong
- rate limited
- connection dropped
- overloaded
- server error
- bad request
- retried
- no agent activity
- answers you marked wrong
Sep 14Sep 20
Sep 20
4requests that failed or were retried
2 failed · 2 retried
8 requests · too few to rate
failed by cause: rate limited 2
1answer you marked wrong
your own entries in the quality log
The total is requests that failed plus requests that were retried, counted from the local transcripts. They are two separate counters and never land on the same request, so nothing here is counted twice. The coloured causes — rate limits, overloads, dropped connections — break down the failed part only: a retry carries no cause of its own. The rate divides that total by the requests the same scan counted, where a request is one response the agent received plus each retry of one — so every retry is counted on both sides and the rate can never pass 100 per 100. It measures how often delivery went wrong per hundred requests, and nothing about whether an answer was right. Only the quality log records that, and only because you wrote it down. Below 10 requests no rate is shown at all, because a ratio taken over a handful of events is noise rather than a measurement — the counts stay, the ratio goes. A day with no agent activity shows —, not 0; a day that ran clean shows 0, because that one was measured. A rate too small to print at one decimal shows <0.1 rather than 0.0, for the same reason: something is not nothing. On your own machine these counts come from your local transcripts.
| Day | Requests that failed or were retried | Requests that failed | Leading cause of the failures | Requests retried | Requests counted (responses plus retries) | Failed or retried per 100 requests | Answers you marked wrong |
|---|---|---|---|---|---|---|---|
| Sep 14 | 8 | 4 | rate limited | 4 | 218 | 3.7 per 100 | 0 |
| Sep 15 | 21 | 11 | rate limited | 10 | 396 | 5.3 per 100 | 1 |
| Sep 16 | 0 | 0 | — | 0 | 168 | 0.0 per 100 | 0 |
| Sep 17 | — no agent activity | — | — | — | — | — | — not recorded |
| Sep 18 | 36 | 19 | rate limited | 17 | 519 | 6.9 per 100 | 2 |
| Sep 19 | 14 | 7 | rate limited | 7 | 348 | 4.0 per 100 | 0 |
| Sep 20 | 4 | 2 | rate limited | 2 | 8 | too few to rate | 1 |
Invented figures, not anyone’s usage — and not a claim about how often AI models are wrong. Your own numbers appear after you connect.
- Counted for you. Failures and retries, counted separately and added up against the requests of the same day — every response the agent got, plus every retry of one — with the failures clustered by cause, so a bad afternoon has a name and a size instead of a feeling.
- Flagged by you. Mark an answer wrong in a click. Only your own entries are ever called that, and they stay in your browser.
- Never guessed. A day you ran no agent reads as a dash, not a zero — silence about a question is not an answer. A day you did run one and nothing went wrong reads as 0, because that one was measured. And a day of only a handful of requests gets its counts and no rate, because a ratio taken over a handful of requests is noise wearing a decimal point.
A small setup. A clearer picture.
Your next session,
with a little more context.
- 01
Connect your computer.
Follow the guided setup to install the local connector. It reads supported agent logs on this machine.
- 02
See your own activity.
Open the dashboard for recorded usage, model costs and cache activity. Missing data is marked as a gap.
- 03
Make it a quick check-in.
Come back after a coding session, or open the live view while you work. Your dashboard is always one click away.
Prefer to look around first? The demo needs no installation.
Private by design
Understand the usage.
Keep the work yours.
Your local connector turns agent logs into usage metrics. Your dashboard gets the counters it needs. Your prompts, conversations, and source code stay on your computer.
Explore the privacy boundaryAgent logsLocal connectorYour dashboard
On your computer
- No prompt or source code uploads
- No provider API key needed for local logs
- Website analytics only if you opt in
Is Usage Radar free?
Yes. All currently available features are free to use. Explore the demo or connect your local agent logs without an account or payment card.
Which tools can I connect?
Usage Radar reads supported local logs from Claude Code, Codex and Gemini CLI. What it can show depends on the usage those tools record on your computer.
Will the estimate match my actual bill?
It may differ. Cost estimates apply model rates to recorded tokens. Subscriptions, discounts and activity outside your local logs can change what you actually pay. Your provider’s invoice remains the authority.
Are my prompts or source code uploaded?
No. The connector parses agent logs locally and returns normalized usage metrics to your browser. Prompt and response text is excluded from its output, and Usage Radar does not upload your source code.
Why might some of my usage be missing?
The connector can only count what is recorded locally. For example, a Codex task running in the cloud may have no local token log. Coverage gaps and unpriced usage are disclosed rather than silently shown as zero.
Free for everyone
Good work deserves
a clear picture.
Start with your own usage. Come back whenever you need a little clarity.
Local logs. Clear estimates. No account required.