Monitoring for Vibe-Coded Apps: Replit, Cursor, Claude Code, Lovable, v0, Gemini, and Codex
Monitoring for AI-generated apps
Replit, Cursor, Claude Code, Lovable, v0, Gemini, Codex. They all generate code you didn't write by hand. That means bugs end up in places you wouldn't think to look.
These apps rarely crash outright. What happens instead: a form stops submitting, a webhook goes quiet, a payment flow breaks after a refactor. The page loads. Something important stopped working. You find out days later, maybe from a customer, maybe from a revenue drop.
This page goes through what tends to break on each platform and how to set up monitoring with Upflag.
How AI-generated apps fail
You ask the AI to refactor a feature. It touches more files than you expected. The page loads, the UI looks right, the obvious stuff works. But a confirmation email stopped sending. Or a Stripe endpoint changed. Or a CSS tweak hid a button on mobile.
Sentry, Datadog, and New Relic are built around crashes and server errors. AI-generated apps don't usually crash. They just quietly stop doing something.
Platform breakdown
Replit
Full-stack web apps, usually Node.js or Python, hosted on Replit.
Replit makes it easy to deploy often, which means more frequent pushes with less review. The AI assistant can change a shared utility and break three routes at once. Background jobs and cron tasks are especially fragile because there's no visible symptom when they stop.
Watch for server-side errors in API routes, background job failures, and uptime. Replit deployments sometimes have cold start problems.
Cursor
Any stack. Cursor is an IDE, so what comes out depends on the project.
The multi-file edit capability is where things get tricky. A refactor that fixes one module can silently break an import somewhere else. Merge conflicts with AI-generated code also produce subtle bugs that pass a quick visual check.
Watch for JavaScript errors after refactors, broken API calls, and anything that spans files edited in the same session.
Claude Code
Also any stack. Works at the terminal level, so it tends to make bigger architectural changes.
Claude Code can scaffold an entire feature in one go. A lot of new code lands at once, and it's generally well-written, but the volume makes review hard. The spots where new code connects to your existing code are where things break.
After a large session, verify your important flows still work end to end.
Lovable
React and Supabase, hosted on Lovable's infrastructure.
You have the least visibility into the generated code here. Supabase edge functions fail without any frontend indication. State management bugs show up as blank screens or spinners that hang. When Lovable restructures components, any Stripe or email integration you set up can break.
Client-side error tracking is the main safety net. You're probably not reading the generated code, so let the browser tell you when something threw.
v0
Next.js with React Server Components, shadcn/ui, Tailwind. Deploys to Vercel.
Server Components are the tricky part. When one throws, the user sees a blank section or an infinite spinner, but nothing shows up in the browser console. Serverless function limits (cold starts, timeouts, payload size) only appear under real traffic. And like everywhere else, integrations break when v0 restructures your server actions.
Gemini
Often Firebase, Cloud Functions, Angular or Flutter.
Firebase Cloud Functions have cold starts and timeouts that generated code doesn't always account for. The bigger blind spot is Firestore security rules. The app works perfectly in development, then fails in production once rules are enforced.
Watch for server-side function errors, auth flows (Firebase Auth specifically), and any Firestore query that hits production rules you didn't write yourself.
Codex
Output varies a lot depending on how it's used.
Codex writes the happy path well and skips edge cases. API integrations sometimes use outdated patterns or deprecated endpoints. Generated tests can mask real bugs by asserting on implementation details instead of behavior.
Watch for error rates on API endpoints, slow responses (Codex occasionally writes inefficient queries), and client-side errors from unhandled cases.
What to actually monitor
Three things:
- Is your app up? Check every 60 seconds, alert if it's not.
- Are there JavaScript errors in the browser? This is the single most useful signal for AI-generated apps. It catches the silent failures that don't show up in server logs.
- Are your important user flows working? Signups, payments, whatever your app needs to do to function.
Setting up Upflag
One script tag. Works regardless of which tool built your app:
<script src="https://cdn.upflag.io/v1/upflag.js" data-key="YOUR_PROJECT_KEY"></script>
Sign up at upflag.io, get your project key, add the tag to your <head>.
You get uptime checks every 60 seconds, client-side error tracking, status pages, and alerts over email, Slack, or SMS. $15/mo flat, no per-event pricing.