See the report
before you install it.
Kadari's client renders your own LLM spend as a single local file — no account, no upload, no talking to us. Generate one now from an illustrative synthetic workload and read every surface of the real thing before you decide to point it at your own log.
Three steps, all of them local.
The client is zero-dependency Python and makes no network call of any kind. We send design partners the wheel directly — install it offline in one line.
Wrap the calls you already make
One line at the call site. Kadari never re-runs the call, never alters it, and carries none of our analysis engine.
create = wrap(client.chat.completions.create,
provider="openai", recorder=rec)Let it write a local log
Append-only JSONL, on a path you control. Nothing is uploaded — there is no network call in the client at all.
rec = LiveRecorder("kadari_capture.jsonl")Render the report
Runs offline against a dated table of published list rates. The output is one self-contained HTML file you can save, print or forward.
kadari analyze kadari_capture.jsonl \ -o report.html
What's real here, and what isn't.
- The renderer
- The page the button opens is the shipped client's own output — same code path as `kadari analyze`, byte for byte.
- The arithmetic
- Token counts multiplied by published list rates, from a dated snapshot the report names on its own header.
- The honesty surfaces
- The unpriced fine-tune, the priced-share qualifier, the estimate labelling — all real behaviour, not staged.
- The traffic
- An invented e-commerce support-triage and catalog workload. No customer's log was used, or ever will be, for a public page.
- The dollar figures
- Whatever they say, they describe the synthetic log. Your numbers come from your log and will look nothing like them.
- Any savings claim
- There isn't one. The report states spend and one arithmetic ceiling marked as an estimate — never a saving.
One detail worth pointing at: the report is computed from token counts and model names only. It never reads your prompts or your model's answers, so nothing you send it can appear on the page it produces.
And it is entirely self-contained: no scripts, no fonts, no images, no requests. Save it and open it on a machine with the network off and it renders identically — which is the property that lets you forward it internally without forwarding anything to us.
It shows you the question.
We measure the answer.
The report can tell you what you spent and, as a clearly-marked arithmetic estimate, how much of that bill sits on work a smaller model in the same family could have taken. What it cannot tell you — what no arithmetic can — is whether that smaller model would have given an acceptable answer on your task. That is the part we measure, against the outputs you already paid for, and it is the only basis on which we will ever show you a number.
Become a design partnerA small cohort, worked closely. We reply personally.