AI-powered accessibility layer for any webpage — six toggles, zero redesign required.
94.8% of the top one million websites fail basic accessibility standards — an average of 51 errors per page, locking out 1.3 billion people who live with a disability.
Retrofitting is expensive. Redesigns take quarters. Most teams ship inaccessible experiences not from indifference but because the tooling makes accessibility a design-phase concern, not a runtime one.
Multimodal AI models treat format as a parameter, not a constraint. accessibility-booster injects six AI-powered transforms at inference time — no redesign, no CMS changes, no new build pipeline.
| Toggle | What it fixes | WCAG criterion | Tech |
|---|---|---|---|
| Alt text | Images with alt="" get AI-generated contextual descriptions |
1.1.1 Non-text Content (A) | Claude Vision |
| Plain language | Grade 16 jargon rewritten to Grade 6 reading level | 3.1.5 Reading Level (AAA) | Claude |
| Audio | Spoken article summary generated on demand | 1.1.1 Non-text Content (A) | Web Speech API |
| Captions | Flat transcript → expressive captions with tone markers | 1.2.2 Captions (A) | Claude |
| Translation | Dutch, Spanish, or French on demand | 3.1.1 Language of Page (A) | Claude |
| Contrast | Text contrast raised from 2.8:1 (failing) to 12.6:1 (WCAG AAA) | 1.4.6 Contrast Enhanced (AAA) | CSS transform |
Each toggle updates a live accessibility score and shows the before/after impact.
The Screen Reader View panel shows the accessibility tree as VoiceOver or NVDA would traverse it:
- Readable nodes — landmarks, headings, interactive controls (via
aria-label), content, live regions,sr-onlywarnings ⊘aria-hidden nodes — decorative elements that screen readers skip- Read aloud — chains
SpeechSynthesisUtterancecalls node by node, highlighting each row as it is spoken
npm install
npm run devCreate a .env file in the project root:
VITE_ANTHROPIC_API_KEY=your_key_here
Without the API key, the demo runs on pre-generated responses so the toggles still work.
- Svelte + Vite
- Tailwind CSS v4
- Anthropic Claude API — alt text, plain language, captions, translation
- Web Speech API — audio summary and screen reader simulation
Traditional software has one output channel. A website is pixels. If you cannot access that channel — because you are blind, have low literacy, use a slow connection, or are in a loud environment — the content is not for you.
Multimodal AI models accept any input and produce any output. Format becomes a response parameter, not a design phase. When accessibility is an architectural property rather than a feature, it is free at inference time.
The 94.8% failure rate is a failure of intent, not technology. The tools already exist.
- The Web Was Never Built for Everyone — AI Is Fixing That — Medium article on the architecture behind this demo
- AI Is Giving 1.3 Billion People Access to the Web — LinkedIn article with production metrics
MIT — see LICENSE
Built by Oleksander Derechei