Accessibility at Agent Selector
We want every person — including people using screen readers, keyboard-only navigation, magnification, high contrast, or reduced motion — to be able to choose, build, and run AI agents here. Accessibility is engineered into the product's code and design tokens, not bolted on: we deliberately use no third-party accessibility overlay or widget.
Our standard
We engineer to WCAG 2.2 Level AA across the whole product in both display themes (dark and light), and track the WCAG 2.1 AA and Section 508 mappings. We do not claim any legal compliance status; we publish what we test and what we find.
Measures we take
- Color is governed by design-token law: every text and control pair is computed mathematically against WCAG thresholds in both themes, and the build fails if any pinned pair regresses.
- Automated accessibility tests run on every code change: an axe-core census of every page family in both themes, keyboard-navigation protocols (skip link, quiz, dialogs, menus, command palette), 400 % zoom reflow to 320 px, text-spacing overrides, forced-colors and reduced-motion checks.
- Every form field uses one shared accessible contract: a real label, associated help and error text, and errors announced to assistive technology.
- Async outcomes (runs, approvals, builds, downloads, sign-in) announce politely to screen readers — one message per outcome, never a flood.
- A first-party Display & accessibility control in the header (theme, text size up to 150 %, high contrast, reduced motion, link underlines) — built on our own code, saved on your device.
- Sensitive agent actions always pause for your explicit approval, costs are shown before every run, and sign-in is passwordless (no memorization or transcription).
Testing status
Continuous automated testing (axe-core with the full WCAG 2.2 AA rule set, keyboard and visual-adaptation protocols, Chromium) gates every release, in both themes, including signed-in surfaces via a fixture harness. Manual screen-reader sessions (NVDA with Chrome on Windows; VoiceOver with Safari on macOS and iOS) follow a published 10-task script and are run by a person; the current round is scheduled, and results are recorded in our public evidence ledger as they complete.
Known limitations
- Payment happens on our payment provider's hosted checkout pages (Square), which we cannot modify. Our hand-off and return pages are in scope and tested.
- Third-party sign-in and app-connection consent screens (Google, connected apps) are controlled by those providers. Passwordless email sign-in is the equal-path alternative.
- AI-generated text varies by run; we control its structure and rendering, and you can always regenerate.
Found a barrier? Tell us.
Reports go straight to the maintainer. We aim to acknowledge within 2 business days and to give you a substantive answer within 10. You don't need to tell us anything about yourself.
If you prefer, you can also pursue formal channels described at ada.gov. This statement was last reviewed on August 18, 2026, based on our evidence ledger; we review it quarterly.