A digital pit stop for your website.

Because it does not get the final say — you do, and neither does the agent that wrote the change.

The Surgeon writes a patch. The Referee then rebuilds your site, re-measures it five times, compares every screenshot, and decides. The Referee cannot see the Surgeon's work — the reasoning is not in its context, so there is no argument to be persuaded by, only numbers. Most patches lose that argument and are thrown away.

What reaches you is a pull request that already survived a build, a re-measurement and a visual diff. It still needs your approval, and nothing merges without it.

Gemini, through Google Vertex AI. Five agents share the model but not the job: Surveyor, Attributor, Surgeon, Referee and Scribe, each with its own tools and its own authority to stop the run.

The measuring is not AI at all — that is Lighthouse and Playwright on fixed hardware. Agents decide what to try and whether it worked; the stopwatch is a stopwatch.

Markup and CSS, plus image binaries. Concretely: a <picture> wrapper around an <img>, width and height attributes, fetchpriority="high", removal of loading="lazy" from an above-the-fold hero, defer on a blocking script, font-display: swap, an image-set() background declaration, and re-encoded WebP files alongside the originals.

It does not touch your application logic, your framework config, your dependencies, or your build pipeline.

No. It opens one only when a patch passed all six criteria, and it reuses the same branch per route — so a second run that improves the same page updates the existing pull request instead of opening another.

Most runs on a healthy site open nothing, which is the correct outcome and shows in the dashboard as a run with no findings.

Because the site is measured twice — once before the patch and once after — with the same routes, the same five samples, and the same machine class. The pull request body carries both numbers.

Improvement is judged on the median of five loads, not a single run, and a metric has to move by at least 10% to count. Anything smaller is inside the noise this kind of measurement produces.

The attempt is withdrawn. No other Core Web Vital may degrade by more than 5%, the rendered layout is compared box by box, the build must still succeed, and the page must not log any console error or failed request it did not log before.

That last criterion exists because a patch once "improved" every metric by breaking a click handler — less JavaScript ran, so the page got faster and stopped working.

Contents read & write (to clone, and to push a redline/* branch), Pull requests write, Checks write, and Metadata read. Nothing else.

Your repository is cloned into a container that exists for one run and is destroyed after it. Your build command runs there, because measuring a site means building it. Screenshots and Lighthouse reports live in a private bucket for 90 days so the dashboard can show them. Nothing is sent to any third party — no analytics, no telemetry.

Anything that builds to static files. Detection needs a build script in package.json — it then runs npm install && npm run build — so Next.js static exports, Astro, Vite, SvelteKit static and Eleventy work without configuration. The output directory is read from your Vite config if you have one, and otherwise guessed from dist, build, out, public, _site.

If none of that fits — a site with no package.json, a different package manager, or an output directory somewhere else — add a redline.config.json with a buildCommand and outputDir and it will use those instead of guessing. Without either, the run stops and tells you so rather than picking something.

Server-rendered apps that cannot be built and served statically are not supported yet.

Three to eight minutes depending on how many routes it measures, starting within seconds of your push. Two runs never race on one repository — the second is skipped rather than queued behind a ten-minute measurement.

Yes. The demo repository is public and holds every pull request Pitstop has opened against it, including the bodies with before-and-after tables, and the branches for attempts that were withdrawn.