Copilot Money

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Do not build this. The bank connections are a licensed relationship, not an API key.

In most countries, reading somebody's account data requires an aggregator with regulatory permission, contracts with the banks, and an ongoing arrangement you cannot enter as an individual. That is the product. The categorisation that learns from your corrections is buildable; the data getting in is not.

WHAT TO DO INSTEAD
- Keep it if the automatic connections are why you use it
- If you would rather your financial history stayed entirely on your own machine, build the version below and accept that you import rather than connect

WHAT YOU CAN BUILD, AND SHOULD

STACK
- Node 20+ with Fastify, SQLite through better-sqlite3, a PWA for the phone
- Everything local, encrypted at rest, never sent anywhere

WHAT TO IMPLEMENT
- Import from your bank's own export — CSV, OFX, QIF — with a mapping step, a dry run, and duplicate detection by a stable import identifier then by date, amount and payee. A double-imported month is the classic disaster
- Amounts in integer cents with a currency. Never a float
- Payee normalisation: the same shop appears as six different strings across a year, and collapsing them is most of what makes categorisation work
- Rules learned from your corrections: a payee pattern to a category, applied as a suggestion the first time and automatically once confirmed. That is the "learns from you" feature, and it is a lookup table with a confidence
- Balances derived from transactions at query time, never stored, or they will eventually disagree with their own rows
- Reconciliation against the bank's stated balance, which is the only thing that tells you the data is right
- Spending by category and month, and net worth over time

THE ONE THING TO GET RIGHT
This is a complete record of your financial life. Encrypt it, keep it local, back it up encrypted off the machine, and put no analytics in it. Not one event.

THE ONE-LINE VERSION
The connection is licensed and the categorisation is a lookup table. Build the second, import by hand, and keep it on your own machine.