Apps never fit, and AI alone failed you too
Every finance app from the App Store is someone else’s idea of your money. The one that fits is the one you build, but AI without a harness hallucinates the numbers.
A reliable, deterministic
ledger for your AI.
OpenLedger is the harness that keeps your AI from hallucinating your money. It turns scattered bank and credit-card statements into a source of truth your AI can build any financial app on.
npm install -g @aquartier/openledgerThen paste one line into your AI. Two steps, about a minute. See the install guide
Works with the AI you already use
Claude Code
Codex
OpenCode
Grok Build
Cursor- PIPI
Why OpenLedger
Local first, PII-masked, security built in
Everything stays on your machine as local bookkeeping. There’s no cloud, no bank login, no external server storing your data. Your financial history lives in a single local database, and is scrubbed (PII-masked) before anything reaches the AI.
Your own private Plaid for your AI
Scattered statements go in, a reliable ledger comes out. The paperwork your bank already sends becomes the source of truth your AI works from, a private data feed you own end to end.
How it works
Give your AI a ledger
One command, one skill file. Your AI gets a deterministic home for every number. No more making them up.
Hand it your statements
Drop in the PDFs your bank sends every month. Each line becomes double-entry bookkeeping, and anything unclear becomes a question, not a guess.
“Ingest my credit cards and bank statements.”
Vibe-code your dream finance app
A budget planner, a spending tracker, a subscription auditor, a retirement planner. Any app you imagine, on numbers that add up.
“Build me a money coach that knows my spending.”
What your AI builds
Every app below is a prompt away. One ledger underneath, so the numbers agree across all of them.
Spending Tracker
Prompt:
“Install OpenLedger and set up my ledger from my statements. Then build me a spending tracker with envelopes for rent, groceries, and fun, a dashboard of where every dollar went, and a Sankey diagram of my money flow each month.”
A private YNAB that already knows where your money goes.
Personal CFO
Prompt:
“Install OpenLedger and set up my ledger from my statements. Then be my CFO: review my month, flag anything unusual on my accounts, and tell me straight what to cut, what to keep, and what to automate.”
An advisor that knows you best and tells no one.
Budget Planner
Prompt:
“Install OpenLedger and set up my ledger from my statements. Then build me a budget planner that charts my what-ifs: rent going up, taking that trip, or saving 20% more every month.”
See your money in ways no app ever shipped.
And your own.
Describe it. It ships tonight.
Whatever you imagine, running on numbers that balance.
Install
Coding Agents (Recommended)
- Claude Code
- Codex
- Cursor
- OpenCode
- PIPI
OpenClaw- Grok Build
Install the CLI
npm install -g @aquartier/openledgerInstalls the
oledcommand.Run OCR locally (optional)
Download typhoon-ocr1.5-2b in LM Studio and start its local server, then hand this to your AI.
Configure OpenLedger to use my local OCR at http://localhost:1234/v1 (model typhoon-ocr1.5-2b), then run oled doctor to confirm.Only scans and photos need OCR; PDFs with a text layer are read directly.
Add the skill
npx skills add phureewat29/openledgerOr run
oled setup. It writes the skill to.agents/skills/, the shared directory most agents read. Pass--dir <path>, e.g.--dir .claude/skillsfor Claude Code.
AI Apps
ChatGPT Work- Claude Cowork
Install the CLI
npm install -g @aquartier/openledgerInstalls the
oledcommand.Run OCR locally (optional)
Download typhoon-ocr1.5-2b in LM Studio and start its local server, then hand this to your AI.
Configure OpenLedger to use my local OCR at http://localhost:1234/v1 (model typhoon-ocr1.5-2b), then run oled doctor to confirm.Only scans and photos need OCR; PDFs with a text layer are read directly.
Paste into your AI chat
Fetch https://cdn.jsdelivr.net/npm/@aquartier/openledger/ skills/SKILL.md and follow it. oled is installed; help me set up my ledger. If your app can’t fetch URLs,
oled setup --print | pbcopycopies the skill itself. Drop it into the first message.
Under the hood
The model reads your statements; the ledger checks the math. Every posting balances, or it doesn’t post.
Double-entry ledger
Every transaction debits one account and credits another, so the same input always produces the same books.
Local SQLite
One file in ~/.oled with 0600 permissions. Your machine is the database.
PII redaction
Names, account numbers, and free text are masked before anything goes to your AI provider. The ledger file itself never leaves.
Reference datasets
Built-in institution data for the US, Japan, Thailand, and China, so your AI recognizes the bank behind every statement.
Questions, not guesses
When a line is ambiguous, OpenLedger surfaces it for you to resolve instead of letting the model pick.
100%
of your ledger stays in one local file
Zero
bank logins or credentials
26B
model params passes the benchmark
OCR
compatible even with small one
Benchmarked with small models
The harness carries the bookkeeping, so the model doesn’t have to be huge. Small, even local, models run the same statement through the same checks, and the books come out identical.
google/gemma-4-26B-A4B-it · 7,249-row banks/cards statements
PASS
statement rows posted
5,126 of 5,126
card payment rows matched
2,108 of 2,108
refund rows matched
15 of 15
balance matched
7 of 7
truthful claims
5 of 5
benchmarking statements read through typhoon-ocr1.5-2b