a11y-mcp

a11y-mcp - Speak Thai. Fill any English web form. Never touch the keyboard. a11y-mcp turns any semantically sound web form into typed tools an AI agent can call - then lets a Thai speaker read, fill, correct, and submit that form entirely by voice, with every confirmation and error spoken back in Thai before anything is sent. THE PROBLEM Two failures, stacked: 1. Agents "use" the web by guessing pixels. They guess which box is which, guess wrong, and when a page rejects a value they just see a red border - with no idea why. 2. Accessibility tooling is built for English. A Thai person facing an English-only visa, university, or airline form is stuck. These forms are well-marked-up and completely unusable if you don't read English - and nobody translates them. Our lead user is a Thai person in front of an English form. Served for free: blind, low-vision, motor-impaired, and elderly users - anyone who can't reliably use a mouse and keyboard. WHAT IT DOES The page exposes typed WebMCP tools (document.modelContext). A GPT-4o-mini brain listens to Thai speech, calls those tools, and the page answers with structured, machine-readable feedback instead of a red border. The agent self-corrects, asks in Thai when unsure, and never submits without a human clicking confirm. HOW IT USES WEBMCP (THE CORE) Four deliberately-scoped tools: - describe_form() - full field structure, types, constraints. - read_form() - the fields to read aloud to someone who can't see the page. - fill_form(values, confirmed?) - validates each value, returns a filled / failed / needs_human split. "failed" is reactive (page rejected a value, so the agent retries); "needs_human" is preemptive (parser was unsure, so nothing is filled and the agent must ask). "confirmed" is the only way a low-confidence field gets filled - after explicit human approval. - submit_form() - a separate, human-gated tool that opens a confirmation overlay and never fires alone. Error text comes from stable Thai templates keyed by an error enum - never model prose, so the same failure sounds identical every time and works with no model call. WHAT MAKES IT NOVEL - Reads the form aloud in Thai - closes the loop: perceive, fill, confirm - not just write. - Preserve-or-translate per field - names stay in Latin script, free-text Thai becomes English ("maa tiao" / vacation), and select values are matched ("Thai" to TH). - Self-corrects from typed errors - the exact argument for WebMCP, shown live. - Asks in Thai and re-opens the mic - a real spoken clarification dialogue. - Honest confidence badge - a badly-marked-up site isn't a failure; the tool says so instead of guessing. HOW WE BUILT IT TypeScript with Vite and the crxjs plugin, Zod for the schema-once tool boundary, an MV3 extension with a MAIN-world content script (document.modelContext plus the brain) and an ISOLATED-world content script (Paxa plus the mic) bridged over postMessage. GPT-4o-mini turns Thai speech into tool calls; Paxa Labs provides Thai text-to-speech (voice "yoyo") and translation; the Web Speech API handles voice input. CHALLENGES The WebMCP spec is unstable - the API had moved from navigator.modelContext to document.modelContext and lost unregisterTool. We instrumented the live Chrome 152 surface to discover the real contract (async registerTool, JSON-string schemas, duplicate-name errors) and rebuilt against it. WHAT'S NEXT Works on sites with reasonable semantics - the confidence badge tells the truth when they don't. Cross-page memory and custom non-input widgets are deliberately out of scope.