Case study
One founder, a team of Claude agents, and a Top 100 finance app in Vietnam
How Nep built Nếp, an automatic expense tracker for Vietnam, going from first commit to App Store submission in 6 days, and reaching 2,300+ users in its first 10 days.
The problem
In Vietnam, most everyday payments go through bank transfers by scanning a VietQR code. People pay dozens of times a week, but their spending is scattered across bank apps, and logging each payment by hand is tedious, so most people give up tracking within days. iOS apps cannot read banking notifications, so automatic tracking has to work differently.
The solution
Nếp makes the payment itself the logging step. You scan a VietQR code inside Nếp. The app reads the amount and recipient, opens your banking app with the transfer ready, and records the expense automatically. Everything stays on the iPhone: no account, no server, no bank login. Recurring recipients get remembered names. It is a personal expense log, not a bank or payment service: Nếp never holds or moves money.
How it was built: a Claude Code agent team
Nep has one human: the founder, who decides what to build and who it is for. The building is done by specialised Claude Code agents, each with a written role, file scope and checklist. Opus plans and reviews; Sonnet writes and tests code.
| Role | What the agent does |
|---|---|
| Planner and architect | Turns founder decisions into specs, challenges the architecture, keeps a decision log |
| BA and UX | User flows, Vietnamese copy, Apple Human Interface Guidelines checks |
| Engineer | Writes SwiftUI code strictly within the files listed in each task ticket |
| Tester and visual QC | Builds, runs the app on the iOS Simulator, takes screenshots, files reproducible bugs |
| Reviewer | Independent review of every diff before the founder sees it |
| Release | Release checklist, versioning, App Store submission notes |
| Researcher and marketing | Competitor and user research, content plans for launch |
| Scribe and chronicler | Handoff notes, changelog and a public build diary |
14 agent roles in total. Every task is a ticket; nothing is called "done" until the build passes and the screenshots have been checked. The founder approves every commit and every App Store submission.
Timeline
- 12 Sep 2026 First commit
- 18 Sep Version 1.0 submitted to the App Store
- 21 Sep Approved and live; version 1.0.1 approved the same day
- First 10 days 2,300+ users, #88 in Finance on the App Store Vietnam
- Early Oct Version 1.0.8: logging without typing (Back Tap and Apple Pay), transfers between accounts
By the numbers
| Calendar time, first commit to today | 26 days (24 active days) |
| Commits | 291 |
| Swift code | about 34,500 lines |
| Versions shipped to App Store Connect | 1.0 to 1.0.8 |
| Team | 1 founder + 14 Claude Code agent roles |
What we learned
- Small tickets win. Each agent gets one ticket with an explicit list of files. Scoped tasks produce reviewable diffs and far fewer regressions.
- Separate the builder from the checker. The agent that writes code never approves it. An independent reviewer and a Simulator tester catch most issues before the founder looks.
- Write decisions down. A running decision log lets a fresh session pick up the work without re-arguing settled choices.
- Keep the human on product and trust. The founder spends time on users, privacy and what not to build. Agents handle execution.
What's next
Our second app, Living Frame, is a family photo widget for Vietnamese families and is being built the same way. Claude Code remains our whole development team: planning, coding, testing and review for both apps.