Websites that reach everyone.
Headline contrast 1.5:1 fails WCAG
Jangkau (Indonesian for reach) is an AI agent that fixes accessibility problems on a web page, then proves every fix before keeping it. It just did that to the headline above.
A batik shop, before and after Jangkau
Same page, same design. A screen reader user could not tell the product photos apart, the form had no labels, and most of the text was too faint to read. Jangkau fixed all 47 failures in 71 seconds.
Counts are failing elements reported by axe-core, the engine behind most accessibility audits.
It fixes one problem at a time, and checks its own work
Most tools stop at a list of problems. Jangkau keeps going until the list is empty, and never trades one problem for another.
Audit
Runs axe-core in a real browser and takes a screenshot of the page.
Fix
Starts with the most severe problem and asks the model for small, exact edits.
Verify
Audits again and compares the page against four checks.
Keep or roll back
A fix that passes is kept. One that fails is undone and retried with the reason.
A fix is kept only when
- the problem it targets actually went down
- no new problem appeared anywhere else
- no image, link, button or text was deleted
- the page still looks like the original
Our first version made the page worse
We traced every run, found why the agent got stuck, and fixed it three times. Failures left on the batik shop, out of 47:
No checks. Fast, but it never knew when a fix broke something.
Strict checks. It rejected correct fixes and quit on half-finished ones.
Keeps partial progress and tolerates problems it will fix later.
Tells the model exactly which element broke, so the retry lands.
On a real open-source template, StartBootstrap Freelancer, v5 took 33 failures to 0 in 40 seconds without moving the layout.
On W3C's own inaccessible demo page, fetched live: 67 to 4. In CI it runs as a GitHub Action that sends each failure to code scanning with its WCAG and EN 301 549 reference.
Every run is traced in neatlogs. v1 trace 002efc98…5371d4, v4 trace 8ad9968f…42835f. Full IDs and run logs are in the repo.
Cheap enough for every small business
$0.06model cost to fix the batik shop page
1 in 6people worldwide live with a disability
June 2025the European Accessibility Act became enforceable
Run it on your own page
git clone https://github.com/PugarHuda/jangkau
cd jangkau && uv venv && uv pip install -r requirements.txt
python -m playwright install chromium
python jangkau.py path/to/your/page.html