Multi-bank CSV ingestion, rule-based categorization, recurring detection, and keyboard-driven triage. A local-first dashboard for discovering patterns in real household spending. Designed for power, built in the browser, owned end-to-end.
Coded it with me
All data stays in the browser
Bank formats unified
Off-the-shelf budgeting apps want bank credentials, push their own category taxonomy, and treat your data as theirs. None of that fits a household that wants to actually explore its own finances: by category, by merchant, by month, with the analytical lens you choose.
I wanted a dashboard built around one question: where is the money going, and what's worth changing? Powerful, private, local. The data never leaves the laptop. The interface gets out of the way.
Pair-coded with Claude. Vite, React, TypeScript. MUI for components. Dexie wrapping IndexedDB. Zustand for filter state. MUI X Charts for the visualization. No server. No cloud.
Refresh the page and your data is still there. Open the app in a different browser and it's empty, because the data has never been anywhere else.
40 seconds of the actual app. Panel collapse, grouping toggles, category tooltip, Forecast, keyboard-driven Sort triage, drill into a month, isolate two categories for discovery, then back to year-to-date.
Or just try it
Launch the live demoLoads in your browser with synthetic demo data preloaded. Nothing leaves your machine. Drop in your own CSVs to play, or clear the tab when you're done.
I expected the chart to be the hard problem. It was the opposite. Different CSV exports tag categories their own way, or not at all. Getting one consistent taxonomy, then making it correctable in seconds, is what made the dashboard trustworthy.
The first version made you fix transactions one at a time with a pencil icon, which is where motivation went to die. The second version is a dedicated triage view that groups every uncategorized transaction by merchant, sorts the queue by dollar impact (biggest first), suggests the most likely category, and lets you commit each decision with one keystroke. Enter accepts the suggestion, 1–9 pick from the grid, ⌘Z undoes. Going from 700 uncategorized rows to zero takes about three minutes, and it feels good.
The Sort view. One decision categorizes every transaction from a merchant at once, and optionally saves a rule so future imports catch the same pattern.
Once recurring charges (rent, utilities, subscriptions, the gym) are detected and pinned, the rest of the spend reads completely differently. You can finally separate the floor (what you owe) from the discretionary (what you choose). That distinction is the entire game.
Forecast tab. Recurring charges separated from variable budgets, each line individually toggleable.
Strongest at sketching tradeoffs and writing the boring code fast. Weakest when it over-engineered (a three-tier abstraction when I wanted a 12-line function). The fix was on me: give more specific direction, push back, redo it.
The working dynamic landed on I direct, Claude implements, I review. Plain, but it's what stuck.
"An afternoon of pair-coding produced what I'd been failing to build on my own."
Uncategorized transactions group by merchant, sort by dollar impact, and surface one card at a time with a suggested category pre-selected. Enter accepts, 1–9 override, ⌘Z undoes. One decision categorizes every transaction from that merchant, and optionally saves a rule for future imports.
Point the app at a folder. Drop in a CSV at the root, or in any subfolder up to five levels deep. The scanner reads it, hashes the content, and dedups against everything already imported. The same statement re-downloaded under a different name gets skipped.
Settings: Watch folder, Demo data, and Demo mode.
The default tooltip sorted categories by stack order, with no total. A custom renderer sorts them by dollar value and adds a Total row. Every hover answers "what dominated this month?" at a glance.
Categories ordered by dollar value, total at the bottom.
The data lives in IndexedDB. "Clear site data" would wipe it. One click exports the whole app state to a JSON file. Drop it back to restore.
A runtime filter for synthetic data. I added it to screenshot the app without exposing real finances. It stayed in for screen-share calls and anyone who clones the repo.
The strapline at the top of every screen reads "Where is the money actually going?" This is the tool that answers it. Visually, locally, and shaped to how my household actually thinks about money.
Bank CSVs in. A stacked-bar visualization out. Recurring detected, budgets forecasted, uncategorized triaged in keystrokes. Designed and pair-coded with Claude across many sessions. The model does the typing. I do the directing.
If you're shipping local-first tools, designing with AI as your engineering partner, or just curious about the workflow, let's talk. The code is on GitHub. MIT-licensed, no warranties, fork freely.