Document accessibility remediation

Repair the structure. Keep the evidence.

AlloFlow turns a document into a reviewable accessibility package: a semantic rebuild, automated findings, a tagged-PDF pathway, and an explicit record of what still needs human judgment. It is designed to support remediation work—not to manufacture a one-click compliance claim.

Get the portable Skill Read the technical white paper

The current U.S. public-sector timeline

The U.S. Department of Justice's Title II web and mobile-app rule uses WCAG 2.1 Level AA as its technical standard. An April 2026 interim final rule extended the dates shown here. Public schools, colleges, universities, libraries, and other state or local entities should confirm how the rule and its exceptions apply to their own content. This is general information, not legal advice. Read the current DOJ guidance.

April 26, 2027Public entities with a population of 50,000 or more
April 26, 2028Smaller public entities and special district governments

Why WCAG exists

Access is the purpose. The deadline is a planning signal.

WCAG gives teams shared, testable criteria for reducing barriers in digital content. Its four principles help make information usable by people with visual, auditory, physical, speech, cognitive, learning, language, and neurological disabilities—and by the assistive technologies they use.

Perceivable

Information has alternatives and structure that people can perceive in different ways.

Operable

People can navigate and use functionality with the input methods available to them.

Understandable

Language, relationships, instructions, and interactions are clear and predictable.

Robust

Content exposes reliable semantics to current and future user agents and assistive tools.

How the pipeline functions

A staged repair loop with separate evidence layers

The application pipeline escalates from deterministic repair to AI-supported diagnosis only where needed. It keeps HTML accessibility evidence separate from PDF structural validation, and it can roll back changes that make the evidence worse.

  1. Assess

    Inspect the source, detect the format, and identify likely structural and content risks.

  2. Extract

    Recover text, layout signals, links, images, and OCR content when a page is image-based.

  3. Rebuild

    Create semantic HTML and apply deterministic fixes before more invasive repair.

  4. Verify

    Combine rubric review with axe-core and IBM Equal Access checks on the exact HTML.

  5. Improve

    Run a bounded keep-best loop; stop on a plateau and revert evidence regressions.

  6. Package

    Produce reviewable HTML, reports, and tagged PDF; optionally validate PDF/UA structure with local veraPDF.

One score cannot prove accessibility. AlloFlow reports the weakest available evidence layer and keeps semantic review, browser checks, PDF structure, and human source comparison distinct.

Choose the right workflow

The Skill and the MCP connector solve different problems

There is no honest single privacy label for every access path. These options make the processing boundary explicit before a document is attached.

Download available

Portable Agent Skill

Upload a PDF to a compatible AI workspace, let the host model create a strict repair plan, and use bundled deterministic scripts to rebuild and verify outputs. No AlloFlow server, Cloudflare document service, remote MCP, or separate model key is used by the scripts.

Choose a Skill package
Versioned release

Local MCP connector

A 31-tool local connector for Claude Desktop that takes PDF, Office, image and scan, markdown, plain-text, CSV, and spreadsheet inputs, with extraction, redaction, validation, and alternate-format workflows. Offline document tools keep document data local. Some setup and export helpers fetch public dependencies without intentionally sending document content; credential checking sends only the key. Gemini-dependent operations may send the full selected document or derived content using the key you configure.

Download and install
Institution pilot code

Remote institutional MCP

A separate fail-closed design for institution-owned Cloudflare infrastructure, opaque job IDs, retention policy, and long-running workflows. It is not a live public document service and is not the default route.

Understand the architecture

What “evidence-backed” means

Each artifact answers a different question

Automated findings are useful precisely because their limits are documented. Human review is still required for meaning, source fidelity, reading order, alternative-text quality, and real-world assistive-technology use.

01

Semantic output

Does the rebuilt document expose headings, lists, tables, links, images, and language in a coherent structure?

02

HTML checks

What do the current rubric review, axe-core, and IBM Equal Access runs find in the exact rendered HTML?

03

PDF structure

Does the exported PDF pass internal preflight and, when available, independent local PDF/UA-1 validation with veraPDF?

Published portable release · v0.2.9

Start with the neutral Skill; use wrappers for their named workflows

These versioned, checksum-listed GitHub Release assets were published August 11, 2026. For ordinary direct use, start with the neutral Skill. Baseline rebuild requires Python 3.9 or newer; tagged PDF requires local Node, Playwright, and Chromium capability; PDF/UA validation requires a separate local Java and veraPDF setup. Source files are never overwritten, and automatic rebuild is blocked for interactive forms, signed documents, certificates, and legal records. Each package includes instructions, scripts, privacy information, licensing, and verification materials. The checksums below match the published release metadata.

Neutral package

Portable Agent Skill

The ordinary direct-use package for compatible agent workspaces that accept a Skill folder or ZIP.

SHA-256 a3af2c179c32f2b018f6c4003cf8ca93d4917671360922d6dc1f517e9e84d399 Download Skill ZIP
Publisher package

OpenAI publisher package

Use for the OpenAI publisher and submission workflow. For ordinary direct use in a compatible agent workspace, choose the neutral Skill ZIP.

SHA-256 05c5c9fa7b7b5ef47ca52e1bc513e5f9eb3ba9b7205e47c408b6140a63b36fa4 Download OpenAI ZIP
Claude Code wrapper

Claude Code plugin package

Use the packaged wrapper intended for Claude Code. This is not the Claude Desktop MCPB connector.

SHA-256 c12cd24bf2a0b543836c6c82b93a5c2d67e3156f3282ebc693d2759388767ea8 Download Claude Code ZIP
Claude Desktop connector

MCPB extension (local-first)

The 31-tool local connector for Claude Desktop: PDF, Office, image and scan, markdown, plain-text, CSV, spreadsheet, and HTML-audit intake with extraction, redaction, validation, and alternate-format workflows. Offline document tools keep processing on your machine; Gemini-dependent operations run only under a key you configure.

SHA-256 published in SHA256SUMS.txt on the release · provenance: gh attestation verify Download MCPB v0.3.5

View release notes and manifest on GitHub

Go deeper

Choose an explanation for your audience

For accessibility and technical teams

Read the architecture, evidence boundaries, privacy model, standards context, and case-study results.

Read the white paper

For educators and evaluators

Explore interactive examples of keyboard, switch, visual, screen-reader, and communication access.

Open the Accessibility Lab

For implementation planning

Compare Gemini Canvas, the Cloudflare browser edition, Desktop, portable Skills, and connector paths.

See ways to use AlloFlow