Structure
Tags, headings and order
Tagged status, marked content, parent tree, role maps, heading sequence, orphaned MCIDs and reading-order mismatches.
Open tool · MCP
Audit a PDF without sending it anywhere.
A free, local-first accessibility auditor for Claude, Cursor and other MCP clients. It reads the PDF’s real structure, returns cited findings, and writes a report you can send to a reviewer.
- 42 automated checks
- 10 manual checks
- MIT licensed
- 0 uploads
No account, no API key, no model call. The document stays on your machine.
What it finds
The defects that make a PDF fail in practice.
The audit checks the PDF object model, not a screenshot. Findings carry a rule ID, severity, standards mapping, evidence where available and a concrete remediation step.
Tables and forms
Relationships that announce
Table headers, Scope/Headers, empty cells, field names, tooltips, duplicate names and tab order.
Text and pages
Scans, fonts and language
Image-only pages, sparse text, embedded fonts, ToUnicode, document language, display title and metadata.
Evidence
A finding a reviewer can use
Each finding carries a rule ID, severity, standards reference, the page and element when they apply, the evidence the parser saw, and a concrete fix. No mystery score.
Report
Self-contained HTML
One file with no external fonts, scripts or tracking. Print-friendly, responsive and accessible in light and dark modes.
MCP
Built for an agent
Audit one file, audit a folder, explain a rule, list checks, compare reports and render HTML — all local tools.
MCP in the loop
The agent calls a tool. The document never leaves the machine.
Ask an MCP client to audit a PDF. The server returns the findings and, when asked, writes a self-contained HTML report. No upload, no provider account, no model spend.
The report
Readable by a person. Useful to a reviewer.
The score is a summary, not the product. The report leads with the highest-impact fixes, shows the evidence, and lists the manual checks that still need a human and a screen reader.
- Severity counts and a document score
- “What to fix first” with remediation steps
- Every automated check, including skipped ones
- Manual checklist and explicit limitations
- Embedded JSON for your own pipeline
What the reviewer sees
55
out of 100 · 2 critical · 2 serious
- “What to fix first”
- Evidence and remediation
- Manual checks and limits
Install
Run it locally in a minute.
From source
git clone https://github.com/Veduis/pdf-a11y-audit.git
cd pdf-a11y-audit
npm install
node bin/cli.js ./2024-annual-report.pdf \
--format html --out report.htmlThe npm package is pending release. The source install is available now.
MCP client
{
"mcpServers": {
"pdf-a11y-audit": {
"command": "node",
"args": ["/absolute/path/to/pdf-a11y-audit/bin/mcp.js"]
}
}
}Works with any MCP client that supports stdio servers. The server never opens a network connection.
The honest boundary
It is an audit, not a certificate.
Automated rules cannot judge colour contrast, alt-text quality, OCR accuracy or whether the reading order makes sense to a person. The report says so and lists the manual checks. Do not publish a conformance claim just because the score passes.
Free by design
No hosted service, no per-document fee, no inference bill, no account. The package is MIT licensed and runs on your machine.
Evidence over theatre
No “AI-powered” score. Each finding is tied to a rule and a standard, with the page when it applies and a fix that a document author can act on.
Need the whole site covered?
Remediate hosts, remediates and maintains the reader.
Use the local audit to find the work. Use Remediate when you want the transcript, tables, translation, embeds and accessibility requests handled for every published PDF.