shikamaru
Provably correct day-count, holiday-calendar, business-day, and payment-schedule calculations. A small, dependency-light TypeScript library and an MCP server, so an AI agent can get the exact date or number instead of guessing.
Why
LLMs are unreliable at date and money math: they pick the wrong day-count convention, forget market holidays, and miscompute accrued interest. shikamaru does it deterministically and proves it against published reference values. Do not let a model guess your interest accrual or your settlement date.
What it does
- Day-count fraction between two dates under seven market conventions: 30/360, 30E/360, 30E/360 ISDA, ACT/360, ACT/365F, ACT/ACT ISDA, ACT/ACT ICMA (with reference periods and stub decomposition).
- Simple accrued interest: notional x rate x day-count fraction.
- Holiday calendars as rules in code, no data feed:
us-federal,nyse,sifma-us,target,uk. - Business-day math: is-business-day, next/previous, ISDA adjustment conventions (following, modified-following, preceding, modified-preceding), T+N settlement.
- Payment schedules: monthly to annual, backward or forward roll, short or long stubs, end-of-month rule, per-period unadjusted and adjusted dates.
- All of it exposed as a library and as an MCP server.
- Every MCP tool is annotated read-only and returns structured output (typed JSON under an output schema) alongside the plain-text result, so an agent client can consume the numbers without parsing text.
