Quincena

Personal finance where every answer is an interface.

Try it in the browser, in Spanish or English. It needs no account and no key: it opens the whole app on Valentina's made-up account, kept in memory, the one in the store's pictures. "Pregúntale a tu plata" on its home asks a scripted agent about that same account, offline, so its answers say the figures the screens say, and what you save there shows on every screen until you leave the example. The web demo is also the showcase for developers: there "Lo que respondió Gemini" replays five sessions Gemini answered for real, and the conversation's settings let Gemini or a key of your own answer instead of the script. On a phone, "Con datos de ejemplo" on the first screen opens the same example, without those developer options: the phone apps are the product.

Ask where September's money went and you get a donut chart, the two things that moved and the five largest payments. Ask whether you can afford Cartagena in December and you get a goal with a slider: drag it and the arrival date recalculates on the device, with no new message from the agent. Ask what subscriptions you have and you get a list with a box on each row and a footer that says what cancelling the ticked ones would save; you cancel them yourself and say so, because the app never cancels anything.

None of those screens is in the source. An agent composes each one at runtime from a catalog of components and sends it as A2UI, with genui rendering them and genui_gen deriving the catalog from the widgets' own constructors.

The account is made up. Valentina is a designer in Medellín, paid on the 15th and the last day of the month, saving for a trip while she pays off a student loan. Every merchant is made up too.

What it shows

A catalog with no second source of truth. Twenty components and six functions, each written as an ordinary Flutter widget or Dart function and annotated. The JSON schema the agent composes against, the builder that turns its JSON back into the widget, and the example data are generated from the constructor. catalog.json is the same catalog as an agent outside the app sees it, checked against the Dart on every test run.

Interfaces that keep working after the agent is done. The goal planner's slider writes to the data model, and two catalog functions read from it to say when the goal is reached and whether that is in time. The subscription list is a template repeated over the data, each row's box writes to its own entry, and a function over the whole list totals the savings. A form validates with rules the agent wrote, and the message of the first failing rule shows under the field.

The tooling around it. In the web demo, turn on developer mode in the conversation's settings and the genui_gen inspector sits over the conversation: the component tree the agent built, every data path with what reads it, what a screen reader announces, and the messages that got the screen there. "Copiar la sesión" puts the whole session on the clipboard as a genui_gen trace, with what the person typed into a form redacted, ready to paste into an issue and replay.

Two languages. The app speaks Spanish and English, following the device or the choice in settings. The scripted answers, the prompt Gemini gets, the catalog's own words and every amount and date switch together: $ 4.719.400 and 19 sept in Spanish, $4,719,400 and Sep 19 in English.

Your own accounts

On a phone or the desktop, "Con mis cuentas" sets Quincena up with your own money instead of Valentina's: how you get paid, your accounts in pesos, dollars or crypto, and what comes in and goes out. It all stays in a SQLite database on the device. Totals convert to the currency you pick, with the official TRM for dollars and Binance's prices for crypto.

Automatic capture

Payments reach Quincena without anyone typing them.

• iPhone. The app adds a "Record a transaction" action to Shortcuts. An automation on Wallet hands it each Apple Pay payment, one on Message the bank's texts and, from iOS 27, one on Notification the banks' apps. It runs without opening Quincena.
• Android. With notification access, Quincena keeps the notifications that carry an amount next to a currency and lets the rest go by unread. Security codes are never kept, and any app can be muted.
• Screenshots, receipts and texts. A screenshot, a photo or a PDF of a payment is read on the device, with Vision on the iPhone and the Mac and ML Kit on Android. On Android, share it with Quincena from any app; on the iPhone, pick it in "Por revisar", or make a shortcut with the "Read a receipt" action and turn on "Show in Share Sheet". What is read from an image always waits to be confirmed.
• Anywhere. A message pasted into "Por revisar" is read the same way.

Everything goes through one parser, written for how Colombian banks and wallets word their alerts: $45.900,00 or $45,900.00, the merchant, the card's last four digits, a balance that is not the purchase. A receipt is read by its labels: the figure after "Valor" or "¿Cuánto?", who got it after "Para", the date and time it says, not the clock at the top of the screenshot. A payment seen by Wallet, the bank's push and an SMS becomes one transaction, and one already entered by hand is not suggested again. Confirming a capture teaches Quincena that card's account and that merchant's category. With "Registrar solo lo que esté claro" on, the next one goes straight in, and can be undone.

When an alert only says "Compra POS 4512", the location fills the gap. With it on, Quincena looks up the shops within 80 metres of where the phone was, in OpenStreetMap's data through Photon, and proposes the nearest. The coordinates are the only thing that leaves the device. Android asks for "Allow all the time" for this, since payments arrive while the app is closed; the app says why before asking.

Run it

flutter pub get
flutter run -d chrome

The example account runs offline, on a fixed day with fixed rates and prices, in a database in memory that is built again each time it opens and gone when it is left. In its conversation a scripted agent answers with the same components, bindings and function calls a model sends, and every number in its answers comes from that account, so saving an expense or a goal's plan there changes the next answer and every screen of the example. "Importar extracto" offers a made-up statement to try, and "Leer un pago" a bank's message, so neither needs a file or a message of your own.

On iOS and macOS the plugins, Firebase among them, come in as Swift packages (pubspec.yaml turns Swift Package Manager on for this app), so there is no CocoaPods step.

Talk to Gemini

On an iPhone, an Android phone, a Mac or the web, "Pregúntale a tu plata" asks Gemini about your own accounts, with no key. The questions go through Quincena's Firebase project with Firebase AI Logic, which only answers the real app (App Check: App Attest on Apple devices, Play Integrity on Android, reCAPTCHA Enterprise on the web), to Gemini on Google Cloud's Agent Platform, which does not train on what it is sent. Each person signs in anonymously so the project can count their requests, with a cap of 20 a minute set on the project and 30 questions a day in the app. Only asking counts: saving an expense, a plan or a cancellation an answer brought is done on the device through the same tools, with no question spent and no connection needed, and the conversation stays while the app is open. Gemini gets the numbers from tools that run on the device, the demo's plus every account in its own currency, never the database, and "Qué ve Gemini" in the app says what travels in plain words.

What is set up for production, which limits stop spending and which only warn, and what Google may do with what is sent on each plan is in docs/PRODUCTION.md.

On a phone, the example account never asks Gemini: its conversation is the script, so trying it spends none of the day's questions and sends nothing anywhere. Only the web demo lets Gemini, or a key of your own, answer it.

Gemini gets the catalog through genui's prompt builder, with two of its defaults switched off: the chat preset forbids updateDataModel, which every interactive component here relies on, and tells a model that cannot run code to do arithmetic itself, which a finance app must never let it do. It gets the numbers from tools that ask the account, never from its own head. When a surface it sends fails validation against the catalog, genui reports why and the app sends that back within the same turn, so the person sees the corrected answer rather than the broken one.

A simulator, an emulator or a local web build passes App Check with a debug token registered in the project, read at build time from a file that is never committed:

flutter run --dart-define-from-file=tool/app_check.local.json

Never do that for a build you publish: a web build made with the token defined carries it in its JavaScript.

Recording real sessions

GEMINI_API_KEY=your-key flutter test tool/record

asks Gemini each of the five questions the example's conversation offers, over the same story, and writes what it sent to assets/traces/ as genui_gen traces. The app replays them in "Lo que respondió Gemini", step by step and with no network, and test/recorded_test.dart replays every one against the current catalog on each run, so a catalog change that would break a real conversation fails a test instead of a person.

After changing a widget or a function, regenerate the catalog:

dart run build_runner build --delete-conflicting-outputs

Tests

flutter test

Besides the answers themselves, in both languages and at twice the system text size, the catalog is held to three things. genUiFuzz renders everything the schema allows and fails on anything that breaks. genUiSemanticsAudit fails on a control a screen reader cannot name. And catalog.json has to match the Dart.

The first two earned their place on the first run. The fuzzer found that every component crashed under any theme but the app's own, and the audit found that the goal slider announced its value without saying what the value was of.

Layout

Path	What is there
lib/catalog/	the components the agent composes with, and the data shapes they take
lib/functions/	the functions the agent can call from a binding
lib/agent/	the assembled catalog, the prompt, the tools, the scripted agent and the live one
lib/data/	the sample account: movements, subscriptions, the goal
lib/money/	assets, decimal amounts, exchange rates and where they come from
lib/domain/, lib/store/	accounts, movements and pay schedules, and the SQLite database they live in
lib/capture/	the parser, the deduplicator, the inbox and the lookup of nearby shops
lib/ui/	the conversation, and in own/ the screens for your own accounts
ios/Runner/	the Shortcuts actions, and the text reader the Mac shares
android/app/src/main/kotlin/	the notification listener, the share target and the text reader

Credits

Type is Bricolage Grotesque and Geist, under the SIL Open Font License. Icons are Phosphor, under the MIT license. The licenses sit next to the files in assets/.

License

Quincena is under the MIT license.