Money Tracker

Your money lives in four places, and none of them can tell you what to do next. Money Tracker combines statements from a bank, a multi-currency account, a share broker and a super fund into one report that opens with what to act on. An AI agent runs the monthly update, and a person stays in charge of anything that matters.

See the live demo (fake data) · What it shows · How it works · Try it

Why I built it

My money is spread across CommBank, Wise, CMC Invest and IOOF. Each exports a different format, the same dollar often appears in two of them, and none of them tells you whether a charge looks wrong or what deserves attention. Many budgeting apps also ask you to connect your bank accounts. I wanted something that keeps statements in my own folder and gets the numbers right, then tells me the few things worth doing this month.

What this project shows

• Getting messy data right: reconciling overlapping sources and file formats so each dollar is counted once.
• Putting an AI agent to work safely: deciding what stays in deterministic code, what the agent handles, and what always needs a person, then writing those limits down as rules it follows.
• Designing for a decision: a report that leads with findings and next steps, with the detail one click away.
• Handling sensitive data with care: personal details live outside the code, and an audit checks every public release for them.

Highlights

• Combining sources without double counting. A transfer from the bank to Wise appears in both. It's replaced by what Wise actually paid out and to whom, matched within 5 days with Wise's fee allowed for. Export formats that overlap and spell IDs differently are de-duplicated, and a ledger and newer PDFs that overlap on the same day are joined without repeats.
• An agent-run monthly update with guardrails. A Claude Code skill downloads the new statements in a browser, files them, rebuilds the report and investigates every alert before involving me. I sign in myself; it asks before each download, always logs off, never moves money, and updates its own runbook when a website changes.
• Checks that explain themselves. Duplicate charges, subscription price rises, bill jumps, big first-time merchants, category spikes and balance gaps, each tuned against false alarms (a daily $5.50 coffee is not a duplicate) and each saying why it fired.
• A dashboard that starts with what matters. The top of the report is a few findings, each with one next step, plus this month's quick checks. Charts, accounts and transactions stay folded away until you want them.
• One self-contained report. Hand-drawn SVG charts, light and dark themes, no libraries, no network.
• Tested on realistic fake data. A generator writes every statement format the readers handle, with one of each problem planted; the tests run the whole pipeline on it.

Try it

Needs Python 3.11+.

pip install -r requirements.txt -r requirements-dev.txt
python generate_sample.py      # writes fake statements to sample_data/
python analyze.py              # prints the summary and opens report.html
pytest                         # runs the pipeline end to end on fresh fake data

Without a config.toml, everything uses config.example.toml, which points at the fake data. To use your own statements, copy it to config.toml (git-ignored) and fill in your folder, names and recipients.

How it works

flowchart LR
    subgraph Sources["Statements (in your private folder)"]
        L[CommBank ledger CSV]
        P[CommBank PDFs]
        W[Wise PDF / XLSX / CSV]
        C[CMC CSVs]
        S[IOOF super PDF]
    end
    subgraph Pipeline["analyze.py"]
        R["Readers<br/>commbank.py · wise.py · holdings.py"]
        M["Combine<br/>match transfers, de-duplicate,<br/>apply rules and config"]
        Q["Summaries + insights"]
        K["checks.py<br/>Needs a look"]
    end
    L & P & W & C & S --> R --> M
    M --> Q & K --> H["report.html<br/>(data embedded, offline)"]
    A(["Agent: monthly update skill"]) -. downloads & files .-> Sources
    A -. reruns, investigates alerts .-> Pipeline
    U(["You: sign in, answer questions"]) -.-> A

Combining sources

The interesting part is that the same money shows up in more than one place:

This is the beginning of the README. Read the rest at the original. Read the full README ↗