Ideas, tutorials, and updates from our team.
Error tracking for a small web app comes down to four things: catch what breaks, group the repeats, read the stack in your own code, and keep the minute before the crash. How each one works in the browser, what to throw away, and when you need a full monitoring tool.
Claude's custom connectors and ChatGPT have nowhere to paste a token, so BugDump's MCP server now lets them sign in. How the flow works, and the four decisions behind it: one connection for every project, known clients only, checks on every request, and changes that stay attributed.
Ask an agent for a weekly status report and watch it open every ticket one at a time. It is not being dumb — it is doing exactly what your API forced it to do. The N+1 problem, moved from your ORM to your MCP server, and the three questions that fix it.
We gave every priority a stable key so integrations could rely on it. That key quietly became the value we sent to clients too — so making priorities editable meant finding nine readers, four switch statements that fail silently, and one assumption that was wrong in the plan.
A shared API key is fine while an AI agent can only read. Once it can move tasks and post comments, "who did this?" matters. How to authenticate an MCP server with personal tokens, and what we changed in BugDump.
Your app has a dark mode toggle. The chat bubble, feedback widget, and cookie banner ignore it. Why prefers-color-scheme is the wrong signal for embedded widgets, and the small API that fixes it.
A third-party widget that ships 450 KB to every visitor has to justify it. Ours couldn't. Here's the measurement, the fix, and the thing about defer that turned out to be wrong.
A UAT round produces a spreadsheet with sixty rows, half of them unactionable. The problem is rarely the testers. It's that the reporting channel throws away the evidence at the exact moment it exists.
MCP turns your product into something an AI agent can read. That part is easy. The hard part is designing tools an agent uses correctly on the first try — which turns out to be a different discipline from designing a REST API.
DevTools is open on the wrong machine. Here's how to instrument console, errors, fetch, and XHR so a user's browser can tell you what actually happened — and the traps that make naive implementations break the page they're monitoring.
Every developer knows the exchange. A user says something is broken, you ask which browser, and by the time they answer you've lost an hour. We got tired of it.