AI-Assisted Accessibility Skill

The “Accessible.org Audit” skill is the user-facing distribution of an end-to-end accessibility audit pipeline that runs inside chat clients (Claude Desktop today, Claude Code / Claude.ai / ChatGPT next). A user pastes a URL, walks through verification with AI pre-grading, and gets a defensible, signed ACR (VPAT 2.5) — all without leaving the chat.

This document is the developer-facing reference. End-user docs live at apps/web/content/guides/claude-desktop-skill.md.

Architecture

┌──────────────────┐    JSON-RPC        ┌─────────────────────────────────┐
│  Claude Desktop  │◀──── over HTTPS ───▶│  api.theaccessible.org/api/mcp  │
│  (DXT extension) │   Bearer <api-key>  │  (Hono on .4 + Lambda)          │
└──────────────────┘                     └────────────────┬────────────────┘
                                                          │
                  ┌───────────────────────────────────────┼───────────────────────────────────┐
                  │                                       │                                   │
                  ▼                                       ▼                                   ▼
         ┌──────────────┐                       ┌──────────────────┐               ┌─────────────────┐
         │ pdf.* tools  │                       │ url.* tools       │               │ acr.* tools     │
         │ existing     │                       │ Phase 1           │               │ Phases 3-4      │
         └──────┬───────┘                       └──────┬────────────┘               └────────┬────────┘
                │                                      │                                     │
                │           ┌──── enqueue SQS ─────────┤                                     │
                │           ▼                          │                                     │
                │   ┌──────────────────┐               │                                     │
                │   │  url-fetch       │               │                                     │
                │   │  executor        │ Puppeteer +   │                                     │
                │   │  on .4 / EC2     │ axe-core +    │                                     │
                │   └────────┬─────────┘ cms-detector  │                                     │
                │            │                         │                                     │
                │            ▼                         │                                     │
                │   ┌──────────────────┐               │                                     │
                │   │  KV: url-fetch:  │◀──────────────┤                                     │
                │   │  ${jobId}        │  poll/read    │                                     │
                │   │  R2: url-fetch/  │               │                                     │
                │   │  ${userId}/...   │               │                                     │
                │   └──────────────────┘               │                                     │
                │                                      │                                     │
                │            ┌─────── lazy on first acr.queue call ──────────────────────────┤
                │            ▼                                                               │
                │   ┌──────────────────┐    Anthropic Haiku 4.5    ┌──────────────────────┐  │
                │   │  verification-   │◀───────  pre-grader ─────▶│  Anthropic API       │  │
                │   │  queue +         │                            └──────────────────────┘  │
                │   │  pre-grader      │                                                      │
                │   └────────┬─────────┘                                                      │
                │            │                                                                │
                │            ▼                                                                │
                │   ┌──────────────────────────────────┐                                      │
                │   │  Supabase: human_verification    │◀─────────────────────────────────────┤
                │   │  (audit-trail rows)              │  acr.decide                          │
                │   └────────┬─────────────────────────┘                                      │
                │            │                                                                │
                │            ▼                                                                │
                │   ┌──────────────────┐    weasyprint sidecar    ┌────────────────────────┐ │
                │   │  acr-composer    │───────────────────────▶  │  HTML + PDF in R2:     │ │
                │   │  (rendering)     │                          │  acr/${userId}/${jobId} │ │
                │   └──────────────────┘                          └────────────────────────┘ │
                │                                                                             │
                └─────────────────────────────────────────────────────────────────────────────┘
                                              │
                                              ▼
                                  Presigned download URLs
                                  (10-min TTL, returned to chat)

Key components

Component	Path	Purpose
MCP HTTP transport	`workers/api/src/mcp/transports/mcp-http.ts`	JSON-RPC 2.0 endpoint, bearer auth
Tool registry	`workers/api/src/mcp/registry.ts`	All `pdf.`, `url.`, `acr.*` tools registered here
URL tools	`workers/api/src/mcp/tools/url-remediate.ts`	Phase 1 — scan/status/result/conformance/getHtml/editHtml/rescan
ACR tools	`workers/api/src/mcp/tools/acr.ts`	Phases 3-4 — queue/decide/state/generate
URL HTTP route	`workers/api/src/routes/remediate-v2.ts`	`POST/GET /api/v2/fetch-url[:jobId]`
ACR HTTP route	`workers/api/src/routes/acr.ts`	`/api/v2/acr/:jobId/{queue,decide,state,generate}`
URL fetch executor	`workers/batch/src/url-fetch-executor.ts`	Puppeteer rendering + axe-core + CMS detection at scan completion
Verification queue	`workers/api/src/services/verification-queue.ts`	11 per-criterion artifact extractors
AI pre-grader	`workers/api/src/services/verification-prograder.ts`	Anthropic Haiku 4.5, parallel `Promise.all`, 0.85 confidence threshold
ACR composer	`workers/api/src/services/acr-composer.ts`	Merges scan + decisions, renders HTML, calls weasyprint
CMS detector	`workers/api/src/services/cms-detector.ts`	WordPress / Squarespace / Webflow / Wix / Shopify / Instructure Canvas / Drupal / Ghost

Tool surface (18 tools)

`pdf.*` (existing, pre-Phase 1)

Tool	Purpose
`pdf.convert`	Upload + start PDF → HTML conversion
`pdf.status`	Poll job
`pdf.result`	Get violations
`pdf.conformance`	VPAT/conformance report
`pdf.getHtml`	Fetch the converted HTML
`pdf.editHtml`	Surgical find/replace or full rewrite

`url.*` (Phase 1)

Tool	Purpose
`url.scan(url, wcagLevel?, auth?)`	Enqueue scan; `auth` is `{cookies?, headers?, localStorage?}` for protected sites
`url.status(jobId)`	Poll
`url.result(jobId)`	Full result + `platform: {name, confidence, hints?}`
`url.conformance(jobId)`	VPAT criteria + summary + score + platform
`url.getHtml(jobId, variant?)`	Desktop / dark / mobile HTML
`url.editHtml(jobId, edits\|html)`	Find/replace or full rewrite, stored under `remediated/`. Does NOT modify the live URL.
`url.rescan(jobId)`	Re-enqueue, returns new jobId. Does not replay original auth.

`acr.*` (Phases 3-4)

Tool	Purpose
`acr.queue(jobId)`	List items needing human verification, with AI pre-grades
`acr.decide(jobId, criterionId, verdict, note?)`	Record decision; auto-classifies source as `ai-suggested-human-confirmed` / `ai-suggested-human-overrode` / `human-only`
`acr.state(jobId)`	Overall progress
`acr.generate(jobId, signoff?)`	Two-call protocol — first call returns `{needsSignoff: true, prefilled}`; second call (with signoff) returns `{downloads: {html, pdf?}, summary, warnings?}` with presigned URLs (10-min TTL)

Auth model

MCP transport: Authorization: Bearer <api-key> where <api-key> is issued at pdf.theaccessible.org/settings → API Keys. Rate-limited per key.
Per-tool ownership: Every tool that touches a job calls loadJob() which reads url-fetch:${jobId} from KV and verifies job.userId === ctx.userId. 404 on mismatch (no existence leak).
human_verification table: writes always set decided_by_user_id = ctx.userId from the verified JWT. Reads filter by both job_id AND decided_by_user_id. RLS policies present but currently dead code (service-role bypass) — see migration header for the full disclaimer.
Auto-session forwarding for *.theaccessible.org: the remediate web UI auto-injects the user’s Supabase session via auth.localStorage so headless renders hit authenticated pages without the user pasting cookies. Implemented in apps/remediate/src/app/dashboard/page.tsx (buildSessionAuth).

Pre-grader behavior

Model: claude-haiku-4-5-20251001. Set ANTHROPIC_API_KEY in .env (already present on .4; missing on Lambda — Lambda doesn’t currently serve user MCP traffic).
Lazy: runs on first acr.queue call for a given jobId. Subsequent calls read decisions from DB.
Parallel: all items dispatched via Promise.all. Per-item .catch(() => null) ensures one bad call doesn’t poison the batch. Critical — sequential pre-grading would hit API Gateway’s 30s timeout for full scans.
Auto-decide threshold: confidence >= 0.85 && verdict !== 'partial' → write with source: 'ai-auto', exclude from the queue surfaced to chat.
Tolerant parser: strips ``` fences, locates first JSON object, validates verdict enum + clamps confidence to [0,1]. Returns null on any failure → item stays for human.

Adding a new criterion to the pre-grader

Add an artifact extractor to workers/api/src/services/verification-queue.ts (look at the existing 11 — extract111, extract131, etc.).
Add the criterion ID to SUPPORTED_CRITERIA.
Write a focused question template — keep it under ~150 words for token budget.
Add a positive test in workers/api/src/__tests__/services/verification-queue.test.ts showing the artifact extracts cleanly from a sample DOM.
The pre-grader picks it up automatically on next deploy.

ACR document generation

Composer: workers/api/src/services/acr-composer.ts
Render path: composeAcr merges scan vpat.criteria with human_verification rows (human always wins over AI auto), normalizes verdicts to canonical capitalized labels (Supports / Partially Supports / Does Not Support / Not Applicable), computes the score = supports / verified-applicable, then renders to HTML with sign-off block + per-row audit-trail badges.
PDF: HTML → PDF via the existing weasyprint-generator.ts service (calls the accessible-pdf-weasyprint sidecar on .4). When WEASYPRINT_URL is unset, falls back to HTML-only with a warnings entry.
Storage: acr/{userId}/{jobId}.html and .pdf. Presigned via @aws-sdk/s3-request-presigner (lazy-imported), TTL 600s.
Sign-off prefill: pulls full_name and email from profiles. Organization is intentionally not prefilled today (no column on profiles). Date defaults to scan-completion day.

URL escaping

User-controlled fields (URL, page title, criterion reasoning, human note, sign-off fields) all flow through the shared esc() helper before HTML output. URLs rendered into href attributes are additionally scheme-validated via safeHref() — non-http(s) schemes (e.g., javascript:) degrade to #. Important because ACRs are downloaded and shared with auditors.

CMS detection

workers/api/src/services/cms-detector.ts runs once at scan completion in the executor and persists platform: {name, confidence, hints?} on the KV record. Subsequent url.result / url.conformance calls read from KV — no per-poll R2 reads.

Currently detected: WordPress (with builder hints — Elementor, Divi, etc.), Squarespace, Webflow, Wix, Shopify, Instructure Canvas, Drupal, Ghost. Confidence ∈ 'high' | 'medium' | 'low'.

Detection rules favor multiple weak signals over single strong ones to avoid false positives. The Instructure Canvas detector is a scoring rule (≥3 of 7 weighted signals required) with an explicit guard against bare HTML <canvas> element collisions — see the in-file comments before extending.

Deployment

The MCP server runs on both Lambda (index-aws.ts) and the .4 Node servers (server.ts). User MCP traffic via api.theaccessible.org routes through the Cloudflare tunnel to .4 — Lambda is currently a hot standby for the API surface.

After changes to MCP tools or executor

# Lambda (optional but recommended for parity)
cd workers/api && npm run build:lambda
cd ../../infra/cdk && AWS_PROFILE=accessible npx cdk deploy AccessiblePdfProd-Api --require-approval never

# .4 Node servers (REQUIRED — this is what api.theaccessible.org hits)
npm run rebuild  # from repo root

After Supabase schema changes

Apply via the Supabase Dashboard SQL Editor (the project doesn’t have CLI migrations wired up). Migrations live in supabase/migrations/<timestamp>_<seq>_<name>.sql.

The current schema includes human_verification (Phase 3 migration 20260508_100_human_verification.sql).

Skill bundle

The DXT bundle lives at apps/skills/accessibility/dxt/. Rebuild after manifest or system-prompt changes:

cd apps/skills/accessibility/dxt
npx -y @anthropic-ai/dxt pack
mv dxt.dxt accessible-org-<version>.dxt

(Note: @anthropic-ai/dxt was renamed to @anthropic-ai/mcpb — migrate when convenient.)

The Anthropic Skill at apps/skills/accessibility/skill/SKILL.md doesn’t need packaging; it’s loaded as-is by Claude.ai / Claude Code.

Costs to watch

Operation	Cost	Notes
URL scan	~$0.02–0.05	Existing — Puppeteer + axe-core run, no incremental MCP cost
`acr.queue` first call	~$0.005–0.02	Anthropic Haiku 4.5 × 11–18 criteria (parallel). Subsequent calls free (cached in DB).
`acr.generate`	~$0.001	Composer + R2 writes; PDF generation free (sidecar)

All AI calls log to the standard request logs. Cost tracking per the global standards is not yet wired into a cost_tracking table — open follow-up.

Troubleshooting

`url.scan` succeeded but `acr.queue` returns nothing

The scan didn’t produce any not-verified criteria — either everything was auto-decidable by axe-core, or the page is well-instrumented. Confirm with url.conformance — look at the criteria array and count not-verified entries.

`acr.queue` is slow (>20s) on first call

Pre-grader is running. Should be ~5–10s with parallel Promise.all. If it’s hitting 30s+, check:

ANTHROPIC_API_KEY is set on the host serving the request (verify .4 env via ssh -i ~/.ssh/nightly-audit larry@10.1.1.4 'grep -h ANTHROPIC_API_KEY ~/accessible/workers/api/.env*')
Anthropic API rate limits (look for 429s in the docker logs)

Claude Desktop shows “Server disconnected”

Almost always mcp-remote’s npx cache is corrupted. Fix:

rm -rf ~/.npm/_npx

Then quit + reopen Claude Desktop.

`acr.generate` returns `warnings: ["pdf generation skipped — WEASYPRINT_URL not set"]`

Lambda doesn’t have weasyprint reachable; .4 does. If the request landed on Lambda, only HTML download is available. User can re-run from a host that has weasyprint, or accept HTML-only.

Auth payload didn’t take. Check:

For *.theaccessible.org: the auto-session-forwarding code in apps/remediate/src/app/dashboard/page.tsx should have run. Verify the request body to /api/v2/fetch-url includes auth.localStorage.
For other sites: auth.cookies is subject to Chromium’s 4096-byte per-cookie limit. Use auth.localStorage for larger session blobs (Supabase, etc.).

Phase log

Phase	Shipped	Description
1	2026-05-08	`url.*` MCP tools (PRs #596)
2	2026-05-08	CMS detection coverage + LLM-tuned tool descriptions (#598)
3	2026-05-08	`acr.queue` / `acr.decide` / `acr.state` + AI pre-grader + `human_verification` table (#599)
4	2026-05-08	`acr.generate` with sign-off + signed download URLs (#600)
5	2026-05-09	Skill bundle (DXT + Anthropic Skill formats) (#602)

Open follow-ups

Cost tracking — wire AI calls into a cost_tracking table per global standards
Lambda ANTHROPIC_API_KEY — set for parity with .4 in case routing changes
DXT package migration from @anthropic-ai/dxt → @anthropic-ai/mcpb
Manifest-tools/list lint — diff dxt/manifest.json’s tool list against the live MCP tools/list to catch drift
ACR organization prefill — source from a tenant/team table once one exists
Apply-then-undo on *.editHtml — currently apply-immediately with no undo stack
Two parallel ACR renderers (acr-report-renderer.ts for files, acr-composer.ts for URLs) — extract a shared HTML shell when a third renderer is needed
Phase 6: PDF flow gets the same acr.* treatment URLs got