Open-source AI maps design tokens to accessible alternatives

AI · 5 min read

A community-backed AI model released this week can ingest a design system’s token file and output candidate accessibility-safe variants — higher-contrast color scales, reduced-motion token sets, and spacing that respects focus targets. The project, built on lightweight transformer architectures and trained on public design-token repositories plus accessibility audit data, focuses on suggestions rather than automatic rewrites so teams retain control.

Maintainers can run the model as a local CLI or host it as a microservice that integrates with token registries. It generates contextual change proposals (for example: "increase background-to-text contrast from 3.2:1 to 4.5:1 and suggest alternative accent colors") and produces a mapping file that preserves semantic token names while adding accessible aliases. The output includes rationale and links to relevant WCAG techniques to help designers and engineers evaluate the suggestions.

Early adopters report the biggest win is bringing accessibility into token decisions rather than retrofitting components later. Teams we spoke to used the tool to create 'accessibility token sets' that are toggled automatically for known assistive scenarios (high-contrast mode, reduced motion, enlarged touch targets), and to generate pull requests with suggested changes that human reviewers then refine.