pixelpact

Your coding agent cannot see the page it just built. pixelpact measures it against the reference and hands back numbers.

A coding agent writes the CSS and says it is done. It never sees the result. Nothing hands it a number, so the loop cannot close: the agent guesses, you open the page, you send it back, it guesses again. Every round costs you the one thing the agent was supposed to save.

The same gap exists without an agent, when a teammate says a section is finished and nobody in the room has a measurement. The difference is that a person can at least look.

pixelpact reads the reference page and writes down what it actually renders: sizes, colors, spacing, typography, hover and focus states, animation keyframes, design tokens. That file is the contract. Point pixelpact at the implementation and it answers with one line per property that drifted, which is precisely what an agent can act on.

npx pixelpact extract https://reference.example.com -o contract.json
npx pixelpact check contract.json http://localhost:3000

Works with

If a browser can render it, pixelpact can measure it.

Related MCP server: websight

For coding agents

An agent that writes UI code cannot tell whether it succeeded. pixelpact-mcp gives it the measurement, so the loop closes without a person in the middle: extract the contract once, then let the agent check its own work, read the deviation list, fix, and check again.

// .mcp.json
{
  "mcpServers": {
    "pixelpact": { "command": "npx", "args": ["-y", "pixelpact-mcp"] }
  }
}

Tools exposed: extract_contract, check_implementation, diff_pixels, read_contract_summary.

Why this is not visual regression testing

Percy, Chromatic, Applitools, BackstopJS and Pixeleye all compare your page against a baseline that you approved earlier. That model works once the page already looks right and you want to keep it that way. While you are still building toward a design, there is no baseline to compare with, and the first run of a regression tool simply records whatever you produced.

The two models are complementary. Use a regression tool to keep a finished page finished, and pixelpact to get it finished in the first place.

Problems pixelpact solves

Install

pnpm add -D pixelpact playwright
pnpm exec playwright install chromium

playwright is a peer dependency, so the browser download stays under your control and a project that already has Playwright installs nothing extra. Node 22.12 or newer is required.

Quickstart

1. Extract the contract from the reference.

npx pixelpact extract https://reference.example.com \
  --selector "main" \
  --viewport desktop,mobile \
  --screenshots .pixelpact/shots \
  -o contract.json

2. Measure your implementation.

npx pixelpact check contract.json http://localhost:3000 --viewport desktop

3. Fix what it reports, then run it again. The command exits 0 when everything is inside tolerance and 1 when it is not, so it drops straight into a script or a CI job.

What a contract looks like

A contract is plain JSON, readable and diffable, with no proprietary format and no service behind it.

{
  "version": 1,
  "source": { "type": "url", "value": "https://reference.example.com" },
  "root": "main",
  "extractedAt": "2026-09-06T01:20:44.812Z",
  "viewports": [{ "name": "desktop", "width": 1440, "height": 900 }],
  "tokens": { "--brand-600": "rgb(11, 114, 133)" },
  "keyframes": { "fade-up": [{ "offset": "0%", "css": "opacity: 0" }] },
  "byViewport": {
    "desktop": {
      "documentHeight": 4218,
      "elements": [
        {
          "selector": "main > header > a.cta",
          "tag": "a",
          "text": "Get started",
          "box": { "x": 120, "y": 32, "w": 148, "h": 44 },
          "styles": {
            "background-color": "rgb(11, 114, 133)",
            "font-size": "16px",
            "border-radius": "8px"
          },
          "hover": { "background-color": "rgb(8, 90, 105)" },
          "focus": { "outline": "2px solid rgb(11, 114, 133)" }
        }
      ]
    }
  }
}

Because it is a file, you can commit it, review it in a pull request, hand it to another developer, or hand it to an agent.

What a check prints

pixelpact check  FAILED
  target    http://localhost:4173/impl.html
  reference http://localhost:4173/ref.html
  viewport  desktop 1440x900
  elements  14 matched, 0 missing of 14
  checks    1056 passed, 10 failed (99.1% of 1066)

deviations (10)
SELECTOR          PROPERTY                  EXPECTED                ACTUAL                  DIFF
body > main > h1  font-size                 48px                    44px                    4px
body > main > a   box.width                 117.75px                109.75px                8px
body > main > a   padding-right             24px                    20px                    4px
body > main > a   padding-left              24px                    20px                    4px
body > main > a   background-color          rgb(11, 114, 133)       rgb(37, 99, 235)        60.1 (color)
body > main > a   border-top-left-radius    8px                     4px                     4px
body > main > a   border-top-right-radius   8px                     4px                     4px
body > main > a   border-bottom-left-ra...  8px                     4px                     4px
body > main > a   border-bottom-right-r...  8px                     4px                     4px
body > main > a   focus.outline             rgb(11, 114, 133) s...  rgb(37, 99, 235) so...  differs

That is a real run against two copies of one page with four declarations changed. Four edits produce ten measured deviations, because padding moves the box width and one border-radius shorthand sets four corners. Run the same check against the reference itself and all 1066 assertions pass, which is the property that matters: a passing check has to mean something.

Add --json to get the same report as a data structure, which is what CI jobs and agents read.

What a side by side run shows

check says which values moved. diff says how many pixels moved. Neither tells a person where to look. side splits both pages into sections, puts them next to each other, and boxes what differs.

npx pixelpact side https://reference.example.com http://localhost:3000 --widths 1440,390

#   SECTION   WIDTH   VERDICT  DIFF
01  hero      1440px  PASS     0.000%
02  features  1440px  FAIL     0.675%
  .pixelpact/side/1440/02-features.png
03  pricing   1440px  PASS     0.000%
04  foot      1440px  FAIL     1.265%
  .pixelpact/side/1440/04-foot.png

That is a real run against the two files in examples/side, which are copies of one page where the card gap and the corner radius were changed. Reproduce it with:

npx pixelpact side "file://$PWD/examples/side/reference.html" \
                   "file://$PWD/examples/side/implementation.html" --widths 1440

This is the command to run before telling anyone that a page is finished.

Features

Commands

Shared flags cover the browser context (--viewport, --selector, --wait, --timeout, --locale, --timezone, --channel, --headful), the output (--out, --json, --quiet), and the parts of extraction you may want to cap (--max-elements, --max-states, --mask). Run pixelpact <command> --help for the full list.

Exit codes

Programmatic use

import { extract, check, formatCheckReport, writeContract } from 'pixelpact-core'

const contract = await extract({
  url: 'https://reference.example.com',
  selector: 'main',
  screenshotDir: '.pixelpact/shots',
})
await writeContract('contract.json', contract)

const report = await check(contract, { url: 'http://localhost:3000' })

console.log(formatCheckReport(report, { color: true }))
if (!report.ok) process.exitCode = 1

Everything is typed, and the contract and report shapes are validated at the boundary, so a malformed file fails with a readable message instead of a stack trace.

In CI

The Action measures a preview deployment against the contract committed in the repository and keeps a single pull request comment up to date instead of adding one per push.

- uses: jamalkamaladdin/pixelpact/action@v0
  with:
    contract: contract.json
    url: ${{ steps.preview.outputs.url }}
    viewport: desktop
    tolerance: 1

It needs permissions: pull-requests: write and no secret beyond the automatic GITHUB_TOKEN. Every input, every output and a complete workflow are in action/README.md.

Reading a Figma frame

extract recognises a Figma url and reads the frame through the REST API. No browser is launched for this step.

export FIGMA_TOKEN=figd_...
npx pixelpact extract "https://www.figma.com/design/KEY/Name?node-id=12-345" -o contract.json
npx pixelpact check contract.json http://localhost:3000

A Figma layer has no CSS selector, so a Figma contract binds to your markup through data-contract attributes. Name the element after the layer and matching stops depending on how the design happened to nest its frames:

<a class="btn btn-primary" data-contract="Hero/CTA">Get started</a>

Anything with no match is reported as missing rather than guessed from tag names.

Published styles come across as tokens: a color style as its color, a text style as a css font shorthand such as 600 60px/72px Geist, an effect style as a box shadow. A fill that is a gradient or an image has no single css value, so it is left out and named in the warnings rather than approximated.

How it works

1. Playwright opens the reference at each requested viewport, waits for fonts, network and paint to settle, and dismisses cookie overlays.

2. A single function is evaluated inside the page. It walks the DOM under your root selector and records the computed style of every visible element, plus custom properties, keyframes, and the geometry of each box.

3. Interactive elements are hovered and focused, and only the properties that actually change are stored, so a contract stays small enough to read.

4. Checking repeats step 2 against your implementation, matches elements by data-contract attribute, then by selector, then by tag and text, and compares property by property. Lengths use a pixel tolerance, colors use a perceptual distance, and animations are compared by name and timing rather than by string equality.

Repository layout

FAQ

How does an agent use it? Install pixelpact-mcp, point the MCP client at it, extract the contract once, then let the agent call check_implementation after every edit and read the deviation table it gets back.

How is this different from Percy or Chromatic? They compare your page against a snapshot you approved earlier, which assumes the page is already right. pixelpact compares it against the reference, which is what you actually have while you are still building. The two fit together: pixelpact to get the page correct, a regression tool to keep it that way.

Do I need a baseline? No. The reference is the baseline, and that is the entire point.

My markup does not match the reference structure. Will anything match? Element matching tries the data-contract attribute first, then the selector, then the tag and its text. Put data-contract="hero-cta" on your element and matching stops depending on how the reference happened to nest its divs.

The page has content that changes on every load. Will a check ever pass? Mask it. --mask ".carousel" keeps a region out of the pixel comparison, and --max-elements stops a long page from producing a contract nobody can read.

Which browsers? Chromium through Playwright. Running a matrix across engines is out of scope on purpose: this tool measures agreement with a design, not differences between browsers.

Does a passing check mean the page is correct? It means every value in the contract matched inside tolerance. Elements that exist in your page but not in the reference are not flagged, because the contract only describes what the reference contains. Run diff as well when nothing extra is allowed.

Status

Everything described above is built and every number shown came from a real run: extraction from a live page and from Figma, checking, the pixel diff, the side by side images, the MCP server and the Action.

Version 0.3. Settled enough to use, not settled enough to promise: the contract format will gain fields before 1.0. If a value you need is missing from it, open an issue and say which one, because that is the fastest way for it to appear.

Contributing

Development setup, the commands, and the pull request flow are in CONTRIBUTING.md. Security reports go through SECURITY.md.

License

MIT © Jamal Kamaladdin