Guides · Bugs and AI coding agents
Chrome DevTools MCP vs a recorded bug report
Updated Sep 23, 2026
Chrome DevTools MCP lets an agent drive Chrome live: open pages, click, read the console and network, run a performance trace. A recording like Roll's captures a bug that already happened, in the session where it happened, and hands the agent the timeline afterwards. Use DevTools MCP when the agent can reproduce the bug. Use a recording when it can't.
Side by side
| Chrome DevTools MCP | Recorded bug (Roll) | |
|---|---|---|
| When it captures | Live, while the agent works | When the bug happened, before anyone asks the agent |
| Whose browser | A Chrome the server launches (separate profile), or yours via --autoConnect | Whoever saw the bug: you, a teammate, QA, a customer |
| Agent can click and retry | Yes | No, it reads what was recorded |
| Console + network | Yes, live | Yes, with request and response bodies, time-aligned with clicks |
| Performance traces, Lighthouse, heap snapshots | Yes | No |
| Video for humans | Screenshots and screencasts | Full tab video, plus a share link |
| Intermittent bugs | Only if it happens again while connected | Yes, including the last 2 minutes after the fact with Rewind |
Pick Chrome DevTools MCP when
- The agent can reproduce the bug by itself, usually on localhost.
- You want it to try a fix and check the result in the same loop.
- The job is performance:
performance_start_trace, Lighthouse audits, memory snapshots. A recording doesn't do this.
Pick a recording when
- The bug already happened. You saw it once, it hasn't come back, and you want the evidence before it's gone. Rewind keeps the last 2 minutes so you can save after the fact.
- Someone else saw it. A PM, QA or a customer. They record, send a link, and your agent opens it by URL. Nobody installs a debugging server.
- It depends on real state: a production account, a feature flag, data the agent's clean profile doesn't have.
- You want the order of events: this click, then this request returned 500, then this error, which is what
timeline.mdis.
What about BrowserTools MCP?
BrowserTools MCP (AgentDesk) takes a middle path: a Chrome extension streams console, network and screenshots from your real, open browser to a local MCP server. It shows what's happening now in your own session. Like DevTools MCP it's live, so the same rule applies: it helps while the bug is happening in front of it, not after.
Setting either up
Step-by-step guides for each editor: Cursor and Claude Code. The overview of all the options is in how to give your agent browser logs.
Questions
Does Chrome DevTools MCP use my normal Chrome profile?
Not by default. It launches Chrome with its own profile in ~/.cache/chrome-devtools-mcp/chrome-profile, so you're signed out of everything there. With Chrome 144+ you can enable remote debugging at chrome://inspect and start the server with --autoConnect to attach to your running browser instead.
Can I use both?
Yes, and it's a good loop. Record the bug where it happened, let the agent read the recording to find the cause, then let it use DevTools MCP or its own browser to confirm the fix.
Is a recording just a worse live session?
It's a different thing: evidence. A live session only shows what happens now. A recording is what happened then, which is the only thing you have for an intermittent bug or one a user hit on their account.