Every token,
accounted for.

Token-Spend Toolkit · financial control for LLM spend across Claude and ChatGPT

Every team buying LLM capacity shares the same blind spot: the bill arrives after the fact, aggregated into a console nobody watches. The Token-Spend Toolkit closes that gap mid-month instead of at invoice time.

It is one canonical data layer with seven focused tools on top. Ingest pulls usage from Anthropic and OpenAI through six interchangeable sources (seeded mocks, key-free console exports, paginated live APIs) and validates every row against a strict schema before it touches the store. Above that sit forecast and budget tracking, a spend explorer feeding a self-contained Command Centre dashboard, leave-one-out anomaly detection, a back-tester that proves the forecast is honest, an automated TL;DR briefing and an actions layer that ranks what to do next with dollar impact attached.

The discipline is the point. Derived cost is reconciled against provider-reported cost on every ingest and agrees within $0.000001. Unknown models raise instead of silently costing $0. Bad rows land in a reject ledger rather than vanishing. No analysis reads the wall clock, so identical data always renders byte-identical reports, which is what lets 204 passing tests pin exact figures to the cent. An independent audit surfaced eight defects before sign-off; all eight were fixed, re-gated and covered by new tests. The toolkit publishes its own limits too, including exactly which adapters are verified and which are not.

As a showcase it covers the full arc of product engineering: schema design, provider integration, statistics, SQL hardening, security and accessibility, each decision recorded and each claim backed by evidence rather than assertion.

Built in Python on DuckDB and Parquet. Board-signed v1, delivering on a Monday schedule, reproducible from a fresh checkout with one command.