io.github.Davmunrey/trazum

Answers whether a prompt fits a maximum, and if not, whether optimising it would. maxTokens comes from the budgets block of trazum.config.json, matched against the prompt's path; read it yourself, since this server never opens a file. Without a budget somebody set there is nothing to check and this is not the tool to call. Offline and free: no model is called.

Input and output price per million tokens, context window and cacheable minimum, for every model in the bundled catalogue.

Applies Trazum's deterministic rules to a prompt and returns the shorter text, the token counts either side, what the difference is worth per month, and any advisories. Offline and free: no model is called.

States the month's measured position against every ceiling you pass: monthly budget, per-day limit, per-label limits. Measured from the log text alone, priced record by record, with the denominator on every figure — days measured against days elapsed. The distance line is division on the past, labelled as such, and absent under the seven-day floor, on an over, and on a zero rate: this tool never forecasts. Ceilings the log cannot measure are named with the reason instead of skipped. Pass the log text itself: this server never reads files, and session keys are never shown.

Reads a usage log — one JSON object per line, each with a "model" and the "usage" object the API returned — and says where the money went: the spend split, per label and per model, whether caching paid for itself, and which levers would actually move the bill. These are the provider's own billed token counts, not estimates. Pass the log text itself: this server never reads files. Add "label", "session" and "ts" fields to the records to unlock the per-workload findings, conversation growth, the period the log covers, and whether the cache TTL fits how fast the turns arrive; the session key is grouped by and never shown.

Asks what a good prompt needs and assembles one from your answers. Call it with the answers you have; it returns the next question worth asking, what is still missing, and the prompt once the required answers are in. Nothing is generated — the questions are fixed and the words in the prompt are yours, so the same answers always produce the same text. An answer of null is a decline, which is an answer: it is recorded and the follow-up it would have opened is not asked. The draft carries what it costs (estimated, and it says so), whether it fits a budget you state, and what trazum optimize can still recover from it — which should be nothing. It does not claim the prompt is good: that is a judgement about text nobody has run.

Call this before making a model call, when a budget exists and you do not know whether this call fits inside it. Answers yes, no, or cannot-tell. Where the numbers come from: the ceilings are in trazum.config.json in the working directory (spend.maxUsd, spend.monthlyUsd, the limits block), and the spend so far is in the usage log the host already writes. Read both yourself and pass the figures: this server never opens a file. Without a ceiling somebody set the answer is cannot-tell, never a yes nobody measured. A refusal carries the cheaper ways to make the same call (a smaller model the prompt still fits in, a batch window), each priced for this call and each naming what it assumes. The budget consumed is measured from real billed usage you pass in; the cost of your call is an estimate of something that has not happened, and the answer keeps the two apart and says which the verdict rests on. Nothing is called and nothing is spent to produce this answer.