RenderCheck

Assert what users actually see, not just what is in the DOM. RenderCheck takes a screenshot, reads it with OCR on your machine, and fails your Playwright or Cypress test with a clear message when the rendered text is wrong, too faint to read, or not there at all.

Renamed. Up to 0.4.1 this package was published as @freetextfromimage/rendercheck. From 0.5.0 it is @rendercheck/sdk; the API is unchanged. See the CHANGELOG for migration steps.

• Local only. OCR runs in-process (Tesseract.js, WebAssembly). No API key, no upload, no network calls. The English language data ships in the package.
• Catches what DOM tests cannot: a € where a $ should be, text covered by another element, low-contrast text, text the DOM has but the screen does not show.
• Plays well with CI. No native modules (pure-JS PNG/JPEG decoding), no browser-side code.

Expected to find text "[Total: $14.00]" but OCR read "[Total: €14.00]" (Confidence: 96%)
  Diff: Total: [$]14.00  →  Total: [€]14.00

Install

npm install --save-dev @rendercheck/sdk

Requires Node 18+, and @playwright/test (≥ 1.40) or cypress (≥ 12).

Playwright

import { test } from '@playwright/test';
import { expect } from '@rendercheck/sdk/playwright'; // Playwright's expect + toShowText

test('checkout shows the right total', async ({ page }) => {
  await page.goto('/checkout');

  await expect(page).toShowText('Total: $14.00', {
    minConfidence: 85,      // OCR confidence of the matched words (default 80)
    minContrastRatio: 4.5,  // WCAG contrast of the matched words (off unless set)
  });

  // One element: also compared with its DOM text (innerText).
  await expect(page.locator('#total')).toShowText('Total: $14.00');

  // Only part of the screenshot (CSS pixels from its top-left corner).
  await expect(page).toShowText('Pay now', { region: { x: 120, y: 340, width: 200, height: 40 } });

  await expect(page).not.toShowText('Total: €14.00');
});

Like Playwright's own assertions, toShowText retries until it passes or timeout (default 5000 ms) runs out. Already using expect.extend? Import toShowText from the same path and add it yourself.

Prices and dates in the right locale

await expect(page).toShowCurrency(1234.5, { currency: 'EUR', locale: 'de-DE' }); // expects "1.234,50 €"
await expect(page).toShowDate('2026-03-04', { locale: 'en-GB' });               // expects "04/03/2026"
await expect(page).toShowDate('2026-03-04', { locale: 'en-GB', format: { day: 'numeric', month: 'long', year: 'numeric' } });

The expected text comes from Intl, so each locale's separators, symbol position and date order are applied for you. When the screen shows something else, the failure says what it is: The screen shows USD ($14.00), not GBP (£14.00). or The date is formatted for en-US (03/04/2026), but en-GB should show 04/03/2026. Dates use UTC unless you pass dateTimeZone, so they are the same on every CI machine.

Cypress has the same checks: cy.matchCurrency(1234.5, { currency: 'EUR', locale: 'de-DE' }) and cy.matchDate('2026-03-04', { locale: 'en-GB' }). In other tools, expectedCurrency and expectedDate (from the package root) give you the text to pass to toShowText.

Right-to-left and other scripts. Direction marks and non-breaking spaces are ignored in every comparison, so Intl output and DOM text match what OCR reads. Only English OCR data ships with the package. For Arabic, Hebrew, Japanese and the like, download the .traineddata file and set langPath; the language is then picked from the locale (ar-EG uses ara), or set lang yourself. Arabic locales use Arabic-Indic digits, so a page that shows 14.00 where ١٤٫٠٠ belongs fails, which is usually what you want.

Cypress

// cypress.config.ts
import { defineConfig } from 'cypress';
import { registerRenderCheck } from '@rendercheck/sdk/cypress/plugin';

export default defineConfig({
  e2e: { setupNodeEvents(on) { registerRenderCheck(on); } },
});

// cypress/support/e2e.ts
import '@rendercheck/sdk/cypress';

// in a test
cy.visit('/checkout');
cy.matchText('Total: $14.00', { minConfidence: 85, minContrastRatio: 4.5 });
cy.get('#total').matchText('Total: $14.00');

Tested in a real Cypress (13.17, Electron, Linux): examples/cypress runs 5 specs against the mock checkout page, including the failing cases. Other Cypress versions and browsers are untested. Run it with npm run build && npx cypress@13 run --project examples/cypress (add --browser electron).

Command line and AI agents

Check a screenshot from a terminal, a script or an AI coding agent. Same local, offline OCR.

npx @rendercheck/sdk verify screenshot.png --expect "Total: $14.00" --min-contrast 4.5
npx @rendercheck/sdk read screenshot.png --json

Exit code 0 means pass, 1 the assertion failed, 2 an error. --json prints { pass, failure, read, confidence, box, message }.

MCP server

rendercheck mcp is a Model Context Protocol server over stdio with two tools: verify_text (is this text really visible, legible and readable?) and read_text (what does a user see?).

{
  "mcpServers": {
    "rendercheck": { "command": "npx", "args": ["-y", "@rendercheck/sdk", "mcp"] }
  }
}

Claude Code: claude mcp add rendercheck -- npx -y @rendercheck/sdk mcp.

Set up every agent in a repo in one step:

npx @rendercheck/sdk init            # Claude Code, Codex/AGENTS.md, Cursor, Copilot
npx @rendercheck/sdk init --agents claude,cursor --dry-run

It adds a short instruction ("after changing UI, verify the pixels with verify_text") and the MCP server entry. Re-running is safe: only the marked RenderCheck block is updated.

Runs are counted in ~/.rendercheck/usage.jsonl on your machine only, so rendercheck stats can show them to you. Screenshot text is never recorded and nothing is sent anywhere. Set RENDERCHECK_NO_USAGE_LOG=1 to turn it off.

Core

import { verifyText } from '@rendercheck/sdk/core';

const report = await verifyText({
  image: fs.readFileSync('screenshot.png'), // PNG or JPEG
  expected: 'Total: $14.00',
  options: { minContrastRatio: 4.5 },
});
if (!report.pass) console.error(report.message);

recognize(image, { region }) gives the raw OCR result (words, lines, boxes in image pixels), and analyzeWordContrast the WCAG contrast math, if you want to build your own assertions.

Options

| Option | Default | | | --- | --- | --- | | region | whole screenshot | { x, y, width, height } in screenshot pixels | | minConfidence | 80 | minimum OCR confidence (0-100) of the matched words | | minContrastRatio | off | minimum WCAG contrast ratio, e.g. 4.5 (normal text) or 3 (large text) | | exact | false | the region's text must equal the expected text, not just contain it | | ignoreCase | false | PAY NOW is not Pay now unless you say so | | dom | true | with a Locator, compare with the element's DOM text too | | timeout / interval | 5000 / 250 | Playwright: retry window and pause between attempts (ms) | | fullPage | false | Playwright: capture the whole scrollable page | | lang / langPath | eng / bundled | other Tesseract languages: point langPath at the folder with <lang>.traineddata.gz |

Pull request annotations (GitHub Actions)

When an assertion fails, RenderCheck can leave a record for CI: a JSON file and a copy of the screenshot with the failing text outlined in red. The GitHub Action in action/ turns those into pull request annotations on the failing test line ("Expected Total: $14.00, OCR read Total: €14.00, failing region x=140, y=244, 219×24 px"), a job summary table, and a downloadable report with the outlined screenshots.

env:
  RENDERCHECK_REPORT_DIR: rendercheck-report      # on GitHub Actions this is already the default
steps:
  - run: npx playwright test                       # or cypress run
  - uses: ParthibanRajasekaran/text-from-image/packages/rendercheck/action@<ref>
    if: always()

Records are off outside GitHub Actions unless you set RENDERCHECK_REPORT_DIR (set it to an empty string to switch them off there). A failure is recorded once, after the last retry, and a retry that passes removes it. GitHub annotations attach to source lines, not screen coordinates, so the region is written in the message and drawn on the screenshot. Everything stays in your workflow: no service, no token, no network call.

What it can and cannot do

• OCR is not perfect. It can misread l/I/1 and O/0, and it can skip text. Light text on a coloured background is often missed on a busy page: scope the assertion to the element (page.locator('#pay')) or a region. Very faint text, which is exactly what a contrast check looks for, may not be read at all. Then the assertion still fails, reported as "in the DOM but not visible in the rendered pixels".
• Contrast is measured on pixels (text colour vs the colour around it, by k-means), not from CSS. Gradients, images behind text and anti-aliasing make it an estimate.
• It verifies visible text. It does not check layout, spacing or colour beyond contrast; use it with visual-regression tooling, not instead of it.
• English only out of the box.

Network

No code in this package makes a network request. Tesseract's own cache is switched off, so nothing is written to your project either. (Failure records are the one exception, and only when you turn them on: see above.) To check it yourself on Linux: npm run test:offline runs the tests in a network-less namespace (unshare -rn).

Example

examples/playwright/ is a Playwright suite against a mock checkout page (a local HTML file), including the failing cases: wrong currency, low contrast, text covered by another element.

npx playwright install chromium   # once
npm run test:example

examples/github-actions.yml shows it in CI.

License

MIT

Links

• Documentation: https://getrendercheck.com/developers
• How it stays local (architecture and threat model): https://getrendercheck.com/security
• Issues and feature requests: https://github.com/ParthibanRajasekaran/text-from-image/issues
• Security reports: see SECURITY.md