PDF Agent StackBuilding a PDF-expert AI agent

Read, verify, write — grounded in the spec. Four MCP servers and three skills for AI agents.

pdf-spec-mcp — quotes the specification

Returns clauses, definitions and requirements (shall / should / may) from ISO 32000 and 16 more documents. Never opens the PDF under examination.

pdf-reader-mcp — returns what is in the file

20 tools that return text, tables, the structure tree, signature fields and XMP declarations, with their position on the page. Never decides whether the file conforms.

pdf-verify-mcp — decides whether it is genuine and conformant

Cryptographic signature verification, tamper detection, PAdES level observation, PDF/A & PDF/UA validation (veraPDF). evaluate_policy returns one of four verdicts (trust_and_use / use_with_caution / human_review_required / reject).

pdf-writer-mcp — creates and edits PDFs

Creates PDFs from text / Markdown / tables and edits them with 20 tools (merge, split, annotations, metadata, …). CJK font embedding built in.

pdf-trust skill — audits an incoming PDF

Calls pdf-verify, pdf-reader and pdf-spec in order, checks signatures, tampering and conformance, and returns a Trust Report with a four-value verdict.

pdf-publish skill — verifies before delivering

Writes with pdf-writer, reads back with pdf-reader, scores with pdf-verify (veraPDF). Nothing that fails the quality gate is delivered; the result goes into a Publish Report.

pdf-read skill — reads what you need from a large PDF

Measures the document first, then picks how to read it (structure tree / search then read / render as image), and writes what could not be read into a Read Report.

Quick start

jsonc

{
  "mcpServers": {
    "pdf-reader": {
      "command": "npx",
      "args": ["-y", "@shuji-bonji/pdf-reader-mcp@latest"]
    }
  }
}

Start with pdf-reader alone, then add the other servers as you need them. → Getting Started (env / veraPDF / fonts)