How to debug an app Claude Code built

Claude Code can build a full app in an afternoon — and then a real user finds the bug the tests didn’t. The hard part isn’t fixing it; it’s understanding it: which user action, which component, which of the hundred files the agent touched last week.

Mushi Mushi gives Claude Code that understanding. Users report bugs from inside your app, Mushi diagnoses each one in plain English against your actual repo, and Claude Code reads the queue over MCP — so “check what’s broken and fix it” becomes a one-line instruction.

The loop, end to end

Install the SDK — npx mushi-mushi detects your framework (React, Vue, Svelte, mobile, Node) and wires it up.
Register the MCP server — npx mushi-mushi setup --ide claude (see the MCP quickstart) adds Mushi’s tools to Claude Code: list reports, get fix context, check blast radius, dispatch and submit fixes.
A user shakes their phone or clicks the widget — the report lands with a screenshot and Mushi’s read on severity and cause.
Tell Claude Code to work the queue — it pulls the diagnosis and fix context, writes the patch, and can open a draft PR once GitHub is connected. Fixes become lessons the agent sees on the next change.

What makes this different from error monitoring

Error monitors capture what the code threw. Mushi captures what the user felt — including bugs that never throw: wrong copy, dead links, a flow that silently no-ops. For AI-built apps those are the majority, and they’re invisible to a stack-trace tool.

Set it up


npx mushi-mushi
npx mushi-mushi setup --ide claude

Free tier: 50 diagnoses a month, no card. MIT SDKs, AGPLv3 server, self-hostable with one command.

Next: MCP server quickstart · Incident loop walkthrough · Compare with Sentry