Back to Blog
Chrome DevTools MCP on Praesepe LMS: What a Homepage Audit Showed
I ran Chrome DevTools MCP 1.13.0 against public pages of Praesepe LMS, the Vite and React app I recently reviewed with Modern Web Guidance 1.0. Here is what a headless audit reported, and what I would not claim from it.
Back
Engineering
Table of contents
A few days ago I wrote about what Modern Web Guidance 1.0 changed in Praesepe LMS, a Vite and React app. That work stayed in the repo. It does not load the live site. I wanted a second pass that looks at the running homepage.
Chrome DevTools for agents is Google's official way to give a coding agent a live Chrome session. The npm package, chrome-devtools-mcp, ships as an MCP server and as a CLI. I ran 1.13.0, published on November 11, 2026, using the CLI, not a chat session inside Cursor.
What I actually ran
I installed [email protected] and started the CLI with --headless, --isolated, and --no-usage-statistics. Chrome was 156.0.8078.47, on Node 24, on a Linux machine. CPU and network throttling were left off. The numbers below are from that machine's network path, not from my laptop, and not from a mid-range phone on a slow connection.
I stayed on public pages: https://lms.praesepe.id/, /courses, and /pricing. I did not sign in, submit forms, or open account flows. Traces and Lighthouse scores are for the homepage only. Courses and pricing did render after a short load. I did not trace those two.
Official docs show a Cursor path (Settings, then MCP, then New MCP Server) with this config:
json
{
"mcpServers": {
"chrome-devtools": {
"command": "npx",
"args": ["-y", "chrome-devtools-mcp@latest"]
}
}
}I am citing that from the project's client configuration guide. I did not test the Cursor UI in this run.
Performance traces
The docs suggest a first prompt like "Check the performance of ..." to record a trace. The CLI can do the same thing. On a cold homepage load, with the cache bypassed, lab metrics were:
- Desktop, default window: LCP 677 ms, CLS 0.00
- Mobile viewport 390x844, at 3x, with touch: LCP 669 ms, CLS 0.00
INP did not appear in either trace. That is expected here. I did not click or type, so there was no interaction to measure.
There was no Chrome User Experience Report field data for this page. The tool may send trace URLs to the CrUX API by default. I left that default on, and it still reported no data.
The two LCP times are close, but the element is not the same. On desktop, LCP was the horizontal brand lockup, an SVG at /brand/praesepe-lockup-horizontal.svg. Time to first byte was 258 ms, resource load delay 123 ms, and download only 23 ms. Render delay was 273 ms, the largest subpart. The discovery insight failed two checks: the image is not in the initial HTML (the initiator is the main JavaScript bundle), and fetchpriority=high was not set. It was not lazy-loaded. Estimated metric savings were still listed as none.
On the mobile viewport, that lockup is hidden below the sm breakpoint. LCP became the hero supporting paragraph, which is text, so there was no image fetch to optimize. TTFB was 262 ms. Render delay was 406 ms.
Render-blocking requests showed up on both runs: the IBM Plex stylesheet from Google Fonts, and the app's own CSS. Desktop estimated savings were 0 ms. Mobile estimated 94 ms for FCP and LCP. Third-party transfer on the desktop trace was 46.8 kB from Google Fonts and 10.4 kB from Cloudflare.
None of this is a crisis at 677 ms unthrottled. It is a hint if I later want the desktop LCP image in the first HTML, and a reminder not to treat an unthrottled lab load as a phone test.
Lighthouse, console, and network
The lighthouse_audit tool is explicit: it scores accessibility, SEO, best practices, and agentic browsing. It excludes performance. For performance, the tool reference points you back to performance_start_trace. I followed that split.
On both the desktop and mobile navigation audits, the scores were Accessibility 100, SEO 100, Best Practices 96, and Agentic Browsing 100. Lighthouse 13.6.0 reported one failed audit, inspector-issues. The detail is a Content Security Policy issue for the Cloudflare insights beacon at https://static.cloudflareinsights.com/beacon.min.js. The console said the same thing, and also said the policy is report-only, so the violation was logged and no further action was taken. The request still returned 200. That matches a Report-Only CSP I left for the server side in the Guidance write-up.
The homepage network list had 11 requests: the document, fonts, the main CSS and JS, the insights beacon, two brand SVGs, a session check, a courses listing, and a Cloudflare RUM POST that returned 204. I did not ask the tool for individual request bodies or headers.
What I would not skip
Usage statistics are on by default. I turned them off for this run. Performance traces may send page URLs to the CrUX API unless you pass --no-performance-crux. The network tool can return Cookie and Set-Cookie headers when you inspect a single request. I did not do that here. The official warning is broader: if you connect an agent to a Chrome session where you are already signed in, the agent can see and act on that session.
The useful part was the same shape as the Guidance work: a concrete report from an app I already ship, and a few follow-ups I can choose. The desktop LCP image is discovered late, and the analytics beacon is still arguing with a report-only CSP. The CLI made that visible without me driving DevTools by hand.
Sources
- Chrome Docs: Chrome DevTools for agents
- Chrome Docs: Get started
- Chrome Docs: Configuration
- Chrome Developers: Chrome DevTools MCP server (public preview)
- Chrome Developers: Chrome DevTools for agents 1.0
- GitHub: ChromeDevTools/chrome-devtools-mcp
- Release: chrome-devtools-mcp 1.13.0
- Tool reference
- Client configurations
- npm: chrome-devtools-mcp
Comments
No comments yet
Loading comments...