WebPerf Snippets review Performance budgets with real browser controls

Real Chromium tests of TTFB, CLS, INP, baseline regression and exit states, separating controlled local measurements from field-user Core Web Vitals.

AgentIndex · Published 2026-11-07 · Updated 2026-11-07

WebPerf Snippets exposes web performance scripts through a CLI, MCP and agent skills. Our adoption question was whether a CI budget detects an explicitly introduced regression and passes after recovery. We ran the published package against a controlled local page in real Chromium.

Adding server delay, a late banner and a slow click changed TTFB, CLS and INP respectively, with failing budgets. Removing the server delay restored a passing result. An INP run without input remained unmeasured rather than becoming zero, and its required budget failed.

Fixed package and measurement scope

On November 7, 2026, we installed [email protected]. Selected CLI, budget, interaction, MCP and measurement files matched release commit dc3a919ac50708d6d805b0b0f4fbd38cbd88a7f7. The environment was macOS 26.6.2, Apple M1 Pro, Node.js 26.3.0 ARM64, Playwright 1.65.0 and Chromium 158.0.8127.18. Pinned CLI source

The reproduction fixture records dependencies, fingerprints, full results and page generation. The page used loopback only, without login or third-party resources. We called no model, CrUX API or report endpoint. These are local laboratory observations, not production visitor experience or ranking results.

ControlObserved behavior
Normal TTFBThree runs, median TTFB 3ms; 500ms budget passed
Add 900ms server delayMedian TTFB 906ms; budget and regression threshold failed
Restore normal responseTTFB 7ms; budget passed again
Insert a late 240px bannerCLS rose from 0 to 0.2956 and failed its budget
Click a 260ms busy handlerMeasured INP 304ms exceeded its 200ms budget
INP without inputRemained unmeasured; required budget failed
CLS budget on TTFB-only runExit 2 with a workflow explanation

We also exercised the basic Core Web Vitals workflow for LCP and CLS. TTFB is a loading diagnostic, not a substitute for LCP. A small synthetic navigation is not a complete field-user assessment.

Server delay had a recovery control

The same URL first responded normally, then waited an additional 900ms, then returned to normal. Median TTFB across the two three-run groups moved from single-digit milliseconds to approximately 0.9 seconds. Recovery returned it to single digits.

The 500ms budget produced exit 1 for the delayed page and exit 0 for normal and recovered runs. Supplying the earlier JSON with --baseline, TTFB=300ms and --fail-on-regression also classified the change as a regression. Budget implementation

The report warned that three values per side were insufficient for its statistical comparison. We preserve that warning. The experiment validates a deliberately large, controlled delay, not a general speed benchmark or meaningful differences between tiny timing changes. Real comparisons need representative pages, repeated runs and consistent conditions.

Layout shifts and interaction latency are different tests

The stable layout had no measured shift. Inserting a banner 250ms after load raised CLS past the 0.1 budget. That exercises the selected buffered observation path, not every scrolling or page-lifecycle scenario.

For INP we supplied a JSON object containing an interactions array: click a button, then wait 500ms. The handler spent 260ms busy before updating text. Chromium reported an interaction value over the 200ms budget. This represents that particular click, with machine load and browser scheduling still relevant.

Without input, the snippet returned a tracking state and no INP value. The budget exited 1. A caller must distinguish an unobserved interaction from a measured fast one. Interaction execution

A real MCP measurement

We started the package's stdio MCP server, completed the client handshake and listed nine tools. Calling run_snippet actually measured the local page's TTFB and returned a structured metric. No simulated server response replaced the browser.

Tool inventory does not establish coverage of every operation. We did not exercise persistent page sessions, every snippet, CrUX, external reporting or a model's interpretation of the output. MCP makes measurements available; an agent's explanation still needs review. MCP implementation

Adoption decision

Start with a small set of budgets on a stable page, retaining JSON baselines and conditions before expanding. The CLI requires Playwright and a separate Chromium download. Installing the npm package alone does not provide a browser. Its declared Node minimum is 20.12; our execution covered the newer version and macOS environment above.

We did not test real-user traffic, mobile hardware, authenticated pages, other browsers or a production network. Nothing here establishes a Google ranking improvement. The demonstrated value is a reusable measurement and gating path. For evaluating bundled skill scripts, see the Anthropic Skills review. To recover past troubleshooting sessions, see the ctx history-search review.

For inspecting controls and completing browser flows before measuring performance, see the Playwright MCP forms and session review.