PDF Spec (PDF Agent Stack)

PDF SPEC MCP Server

日本語版 README はこちら

An MCP (Model Context Protocol) server that provides structured access to ISO 32000 (PDF) specification documents. Enables LLMs to navigate, search, and analyze PDF specifications through well-defined tools.

[!IMPORTANT] This is a specification reference, not a rule engine. It retrieves and structures the text of ISO 32000 — clauses, tables, definitions, and shall/should/may requirements. It does not examine a PDF file, and it cannot tell you whether a document conforms to anything. Conformance verdicts come from pdf-verify-mcp (validate_conformance / evaluate_policy).

The distinction matters because three different things get conflated: declaration — a label the file wrote about itself ("I am PDF/A" in the metadata). Writing it is not evidence / conformance — whether the file actually meets the standard. There is no way to prove it in full; you can only find where it breaks the rules / validation — what a validator (veraPDF and the like) reports against the checks it implements. A pass means "this inspection did not fail", not "the file conforms to the standard". Reading a shall here tells you what the standard requires — not whether your file meets it.

A search that returns nothing means "cannot answer", not "no such requirement." ISO 19005 (PDF/A) and ETSI PAdES are outside this corpus; see list_specs → coverage.gaps.

What each PDF family server does — and does not do

[!IMPORTANT] PDF specification files are NOT included in this package. You must obtain the PDF specification documents separately and place them in a local directory.

Download from: PDF Association — Sponsored Standards

See "Setup" for details.

Features

• Multi-spec support — Auto-discovers and manages up to 17 PDF-related documents (ISO 32000-2, PDF/UA, Tagged PDF guides, etc.)
• Structured content extraction — Headings, paragraphs, lists, tables, and notes from any section
• Full-text search — Keyword search with section-aware context snippets
• Requirements extraction — Extracts normative language (shall / must / may) per ISO conventions
• Definitions lookup — Term definitions from the "Terms and definitions" clause (clause 3 of ISO 32000-2, clause 4 of ISO 32000-1)
• Table extraction — Multi-page table detection with header merging
• Version comparison — Diff PDF 1.7 vs PDF 2.0 section structures
• Bounded-concurrency processing — Parallel page processing for large documents
• On-disk index cache — The search index and the full requirements scan are built once per PDF and reused by every later process (ISO 32000-2: ~6 s → ~0.2 s)

Architecture

graph LR
    subgraph Client["MCP Client"]
        LLM["LLM<br/>(Claude, etc.)"]
    end

    subgraph Server["PDF Spec MCP Server"]
        direction TB
        MCP["MCP Server<br/>index.ts"]

        subgraph Tools["Tools Layer"]
            direction LR
            T1["list_specs"]
            T2["get_structure"]
            T3["get_section"]
            T4["search_spec"]
            T5["get_requirements"]
            T6["get_definitions"]
            T7["get_tables"]
            T8["compare_versions"]
        end

        subgraph Services["Services Layer"]
            direction LR
            REG["Registry<br/>Auto-discovery"]
            LOADER["Loader<br/>LRU Cache"]
            SVC["PDFService<br/>Orchestration"]
            CMP["CompareService<br/>Version Diff"]
        end

        subgraph Extractors["Extractors"]
            direction LR
            OUTLINE["OutlineResolver<br/>TOC & Section Index"]
            CONTENT["ContentExtractor<br/>Structured Extraction"]
            SEARCH["SearchIndex<br/>Full-text Search"]
            REQ["RequirementExtractor"]
            DEF["DefinitionExtractor"]
        end

        subgraph Utils["Utils"]
            direction LR
            CACHE["LRU Cache"]
            CONC["Concurrency"]
            VALID["Validation"]
        end
    end

    subgraph PDFs["PDF Spec Files (obtained separately)"]
        direction LR
        PDF1["ISO 32000-2<br/>(PDF 2.0)"]
        PDF2["ISO 32000-1<br/>(PDF 1.7)"]
        PDF3["TS 32001–32005<br/>PDF/UA, etc."]
    end

    LLM <-->|"stdio / JSON-RPC"| MCP
    MCP --> Tools
    Tools --> Services
    Services --> Extractors
    Services --> Utils
    LOADER --> PDFs
    REG -->|"Filename pattern<br/>auto-discovery"| PDFs

    style Client fill:#e8f4f8,stroke:#2196F3
    style PDFs fill:#fff3e0,stroke:#FF9800
    style Tools fill:#e8f5e9,stroke:#4CAF50
    style Services fill:#f3e5f5,stroke:#9C27B0
    style Extractors fill:#fce4ec,stroke:#E91E63
    style Utils fill:#f5f5f5,stroke:#9E9E9E

Layer Overview

Setup

1. Obtain PDF Specification Files

[!WARNING] PDF specifications are copyrighted documents and are not included in this package. Download them from the sources below and place them in a local directory.

All 17 files below are supported. You do not need all of them — place only the specs you need (at minimum, ISO 32000-2 is recommended).

pdf-specs/
│
│ ── Standards ─────────────────────────────
├── ISO_32000-2_sponsored_EC3.pdf          # iso32000-2  : PDF 2.0 EC3 (recommended; falls back to -ec2.pdf)
├── ISO_32000-2-2020_sponsored.pdf         # iso32000-2-2020 : PDF 2.0 original
├── PDF32000_2008.pdf                      # pdf17       : PDF 1.7 (for version comparison)
├── pdfreference1.7old.pdf                 # pdf17old    : Adobe PDF Reference 1.7
│
│ ── Technical Specifications (TS) ─────────
├── ISO_TS_32001-2022_sponsored_EC3.pdf    # ts32001     : Hash extensions (SHA-3)
├── ISO_TS_32002-2022_sponsored_EC3.pdf    # ts32002     : Digital signature extensions (ECC/PAdES)
├── ISO_TS_32003-2023_sponsored.pdf        # ts32003     : AES-GCM encryption
├── ISO-TS-32004-2024_sponsored.pdf        # ts32004     : Integrity protection
├── ISO-TS-32005-2023-sponsored.pdf        # ts32005     : Namespace mapping
│
│ ── PDF/UA (Accessibility) ────────────────
├── ISO-14289-1-2014-sponsored.pdf         # pdfua1      : PDF/UA-1
├── ISO-14289-2-2024-sponsored.pdf         # pdfua2      : PDF/UA-2
│
│ ── Guides ────────────────────────────────
├── Tagged-PDF-Best-Practice-Guide.pdf     # tagged-bpg  : Tagged PDF Best Practice
├── Well-Tagged-PDF-WTPDF-1.0.pdf          # wtpdf       : Well-Tagged PDF
├── PDF-Declarations.pdf                   # declarations: PDF Declarations
│
│ ── Application Notes ─────────────────────
├── PDF20_AN001-BPC.pdf                    # an001       : Black Point Compensation
├── PDF20_AN002-AF.pdf                     # an002       : Associated Files
└── PDF20_AN003-ObjectMetadataLocations.pdf # an003      : Object Metadata

2. Install

This package ships a CLI binary (pdf-spec-mcp) intended to be launched by an MCP client. You do not need to install it manually — just point your MCP client to npx @shuji-bonji/pdf-spec-mcp@latest as shown in the next step.

If you want to run it directly from the shell (e.g. for debugging):

PDF_SPEC_DIR=/path/to/pdf-specs npx -y @shuji-bonji/pdf-spec-mcp@latest

Or install it globally (optional):

npm install -g @shuji-bonji/pdf-spec-mcp
PDF_SPEC_DIR=/path/to/pdf-specs pdf-spec-mcp

3. Configure MCP Client

Environment Variable

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "pdf-spec": {
      "command": "npx",
      "args": ["-y", "@shuji-bonji/pdf-spec-mcp@latest"],
      "env": {
        "PDF_SPEC_DIR": "/path/to/pdf-specs"
      }
    }
  }
}

[!IMPORTANT] Use @latest (or pin a version). npx -y <pkg> without a version keeps running whatever it cached the first time — -y only skips the install prompt, it does not check for updates. A bare specifier will happily run a months-old release. @latest makes npx check the registry on each start; pin @0.4.0 instead if you want reproducibility. To clear a stale cache: rm -rf ~/.npm/_npx.

Cursor / VS Code

Add to .cursor/mcp.json or VS Code MCP settings:

{
  "mcpServers": {
    "pdf-spec": {
      "command": "npx",
      "args": ["-y", "@shuji-bonji/pdf-spec-mcp@latest"],
      "env": {
        "PDF_SPEC_DIR": "/path/to/pdf-specs"
      }
    }
  }
}

Installed as a Claude Code plugin

The plugin manifest (.claude-plugin/plugin.json) declares "PDF_SPEC_DIR": "${PDF_SPEC_DIR}" and relies on the host to substitute the reference. Not every host does:

When the reference is passed through verbatim, every tool answers with REGISTRY_ERROR whose message starts with PDF_SPEC_DIR is the unexpanded text "${PDF_SPEC_DIR}" and whose first next_actions entry is set_absolute_path_in_host_config. That is a host-configuration problem, not a missing directory — write the absolute path into the host's own configuration and restart. Host-by-host steps: https://shuji-bonji.github.io/pdf-agent-stack/reference/env-vars.

Index cache

Two operations walk every page of a specification: the first search_spec on a spec builds its full-text index (ISO 32000-2, 1023 pages: about 6 s on a laptop), and get_requirements without a section scans every section (about 11 s). Everything else opens only the pages it needs and answers in well under a second.

Since 0.5.0 those two results are written to disk after the first build and read back by every later process — an MCP client that starts one server per session no longer pays the build each time. The second process answers the same search_spec in about 0.2 s and the full requirements scan in about 0.02 s, from byte-for-byte the same index.

• Location: ${PDF_SPEC_CACHE_DIR:-${XDG_CACHE_HOME:-~/.cache}/pdf-spec-mcp}/v1/<version>/<spec>.<kind>.<sha256[0:16]>.json. The whole 17-spec corpus is about 18 MB per package version.
• Key: package version, pdfjs-dist version, spec id, and the SHA-256 of the PDF. A replaced PDF, an upgraded server, or an upgraded pdfjs all miss and rebuild. Entries of older versions are left in place (another install may still use them); --clear-cache removes everything.
• Failure is a miss, never an error: an unreadable, truncated, or foreign file is rebuilt; an unwritable directory is reported once on stderr and the server carries on without a cache.
• It is derived from your copy of the PDFs and stays on your machine. It is not part of the package and must not be redistributed — the specifications are copyrighted.

Nothing about searching changes: the same in-memory structure is searched by the same code. Only where it comes from (built vs. read) does.

Pre-building the cache

The cache fills lazily, one spec at a time as tools touch it. To warm every spec up front — after installing, after upgrading, or from cron — run the CLI (it uses the same code path as the tools, processes specs sequentially, and exits):

PDF_SPEC_DIR=/path/to/pdf-specs npx -y @shuji-bonji/pdf-spec-mcp@latest --build-cache
#   --spec=iso32000-2,pdf17   only these specs
#   --force                   rebuild even when a valid entry exists
npx -y @shuji-bonji/pdf-spec-mcp@latest --cache-info     # directory, key, entries
npx -y @shuji-bonji/pdf-spec-mcp@latest --clear-cache    # remove the directory

A full build of the 17-spec corpus takes about a minute on a laptop.

Available Tools

All tools accept an optional spec parameter to target a specific specification (default: iso32000-2).

list_specs — Discover Specifications

List all available specification documents. Use the returned IDs as the spec parameter in other tools.

// List all specs
{ }

// Filter by category
{ "category": "ts" }        // Technical specs only
{ "category": "pdfua" }     // PDF/UA only
{ "category": "guide" }     // Guide documents only

get_structure — Table of Contents

Get the section hierarchy (TOC tree) of a specification.

// PDF 2.0 top-level sections only
{ "max_depth": 1 }

// Expand to 2 levels
{ "max_depth": 2 }

// TS 32002 (Digital Signatures) full structure
{ "spec": "ts32002" }

// PDF/UA-2 structure
{ "spec": "pdfua2", "max_depth": 2 }

get_section — Section Content

Get structured content (headings, paragraphs, lists, tables, notes) of a specific section.

A parent section returns its entire subtree (its preamble followed by all subsections, in document order). Top-level clauses can be very large — prefer the most specific section number.

Every answer carries pageRange and boundary. boundary says how each end of the content was cut:

Anything other than heading / heading, heading / page or heading / document means the cut fell on a page edge: read the neighbouring section too before relying on the content. For a parent section, start is the parent's own and end is its last descendant's.

// PDF 2.0 Section 7.3.4.2 (Literal Strings)
{ "section": "7.3.4.2" }

// PDF 2.0 Annex A
{ "section": "Annex A" }

// TS 32002 Section 5
{ "spec": "ts32002", "section": "5" }

// PDF/UA-2 Section 8 (Tagged PDF)
{ "spec": "pdfua2", "section": "8" }

search_spec — Full-text Search

Search across a specification with section-aware context snippets. The first call on a spec builds its index (a few seconds); the index is then cached on disk (see Index cache).

// Search PDF 2.0 for "digital signature"
{ "query": "digital signature" }

// Limit results
{ "query": "font", "max_results": 5 }

// Search within TS 32002
{ "spec": "ts32002", "query": "CMS" }

get_requirements — Normative Requirements

Extract normative requirements (shall / must / may) per ISO conventions.

sectionsCutByPage lists the sections in scope whose content was cut at a page edge (see boundary under get_section); their requirements may include a neighbour's sentences or miss the tail. It is empty when every section in scope was cut at headings.

// All requirements in section 12.8
{ "section": "12.8" }

// Only "shall" requirements
{ "section": "12.8", "level": "shall" }

// Only "shall not" requirements
{ "section": "7.3", "level": "shall not" }

// PDF/UA-2 requirements
{ "spec": "pdfua2", "section": "8", "level": "shall" }

get_definitions — Term Definitions

Look up term definitions from the "Terms and definitions" clause — clause 3 of ISO 32000-2 (71 terms), clause 4 of ISO 32000-1 / pdf17 (48 terms). section in each result is the number printed in the document (3.12, 4.11).

// Search for "font" definitions
{ "term": "font" }

// List all definitions
{ }

// PDF/UA definitions
{ "spec": "pdfua2", "term": "artifact" }

get_tables — Table Extraction

Extract table structures (headers, rows, captions) from a section. Multi-page tables are automatically merged.

The answer carries pageRange (the pages the tables were read from) and boundary (see get_section). totalTables: 0 means no Table element was found in those pages — with a page-cut boundary, the pages themselves may not be the whole section.

// All tables in section 7.3.4.2 (Table 3 — Escape sequences)
{ "section": "7.3.4.2" }

// Specific table only (0-based index)
{ "section": "7.3.4.2", "table_index": 0 }

// TS spec tables
{ "spec": "ts32002", "section": "5" }

compare_versions — Version Comparison

Compare section structures between PDF 1.7 (ISO 32000-1) and PDF 2.0 (ISO 32000-2). Uses title-based automatic matching to detect matched, added, and removed sections.

[!NOTE] This tool requires both PDF 1.7 (PDF32000_2008.pdf) and PDF 2.0 files in PDF_SPEC_DIR.

// Diff section 12.8 (Digital Signatures)
{ "section": "12.8" }

// Compare all top-level sections
{ }

Supported Specifications

The server auto-discovers PDF files in PDF_SPEC_DIR by filename pattern matching:

Directory Structure

src/
├── index.ts              # Entry point: MCP server on stdio, or the cache CLI
├── cli.ts                # --build-cache / --clear-cache / --cache-info
├── config.ts             # Configuration & spec patterns
├── errors.ts             # Error hierarchy (PDFSpecError → sub-classes)
├── services/
│   ├── pdf-registry.ts       # Auto-discovery of PDF files
│   ├── pdf-loader.ts         # PDF loading with LRU cache
│   ├── pdf-service.ts        # Orchestration layer
│   ├── index-store.ts        # On-disk cache for the search / requirements indexes
│   ├── compare-service.ts    # Version comparison
│   ├── outline-resolver.ts   # Section index builder
│   ├── content-extractor.ts  # Structured content extraction
│   ├── search-index.ts       # Full-text search index
│   ├── requirement-extractor.ts
│   └── definition-extractor.ts
├── tools/
│   ├── definitions.ts    # MCP tool schemas
│   └── handlers.ts       # Tool implementations
├── types/
│   └── index.ts          # Shared type definitions
└── utils/
    ├── concurrency.ts    # mapConcurrent (bounded Promise.all)
    ├── text.ts           # Text normalization
    ├── cache.ts          # LRU cache
    ├── file-hash.ts      # SHA-256 of a PDF (index cache key)
    ├── validation.ts     # Input validation
    └── logger.ts         # Structured logger

Development

git clone https://github.com/shuji-bonji/pdf-spec-mcp.git
cd pdf-spec-mcp
npm install
npm run build

# Unit tests
npm run test

# E2E tests (requires PDF files in ./pdf-spec/)
npm run test:e2e

# Lint & format
npm run lint
npm run format:check

License

MIT