Monitoring for Claude Code apps
Monitoring for Claude Code apps
Claude Code works at the terminal level. It can scaffold an entire feature in one session: routes, components, database migrations, all of it. The code is generally solid, but the volume makes review hard. You come out of a session with 20 files changed and a vague sense that everything works.
A week later you find out the contact form has been broken since that session.
The failure modes
The spots where new code connects to your existing code are where things go wrong. Claude Code doesn't always know about your auth middleware, your webhook endpoints, or your edge cases. It writes code that works in isolation but breaks a dependency you didn't mention.
Background jobs are a blind spot. If Claude Code rewrites a cron handler or a queue worker, there's no visible symptom when it stops running. You find out when a report didn't send or data looks stale.
And because Claude Code can make big architectural changes in one go, the blast radius of a mistake is larger. A route rename that breaks three downstream consumers. A database migration that changes a column type and quietly breaks a query somewhere else. These don't crash the app. They just make it wrong.
Setup
Run this in your project:
npx upflag init
Claude Code can do this for you. Just ask it to "add upflag monitoring" and point it at the npx command. It detects your framework and adds the snippet to the right file.
Or add two script tags manually to your <head>:
<script>window.__ufq=[];window.addEventListener("error",function(e){window.__ufq.push(e)});window.addEventListener("unhandledrejection",function(e){window.__ufq.push(e)})</script>
<script src="https://upflag.io/api/v1/snippet?key=uf_YOUR_KEY" defer></script>
Sign up at upflag.io for your project key. Errors show up in your dashboard right away.
MCP server
Add Upflag as an MCP server and Claude Code can see your production errors while it's working. Tell it "check upflag for recent errors and fix them" and it pulls the error data directly: which page threw, what the error was, how many users hit it. It goes straight to fixing the problem instead of you having to describe what's broken.