PDF Reader (PDF Agent Stack)

PDF Reader MCP Server

[npm version] [License: MIT] [Built with Claude Code]

English | 日本語

An MCP (Model Context Protocol) server specialized in deciphering PDF internal structures.

While typical PDF MCP servers are thin wrappers for text extraction, this project focuses on reading and analyzing the internal structure of PDF documents. Pair it with pdf-spec-mcp for specification-aware structural analysis and validation.

PDF family

pdf-reader-mcp inspects signature structure (inspect_signatures); for cryptographic signature verification, trust/revocation evaluation, and PDF/A conformance validation, use pdf-verify-mcp.

Features

18 tools organized into three tiers:

Tier 1: Basic Operations

Tier 2: Structure Inspection

Tier 3: Comparison

validate_tagged and validate_metadata were removed (deprecated since 0.7.0). This server reports what a PDF contains and does not return pass/fail against a standard. For PDF/UA or PDF/A conformance, use pdf-verify-mcp validate_conformance (flavour: "pdfua-1", "pdfa-*", …). For the facts those tools used to look at, use inspect_tags (structure tree) and get_metadata (metadata fields) here.

read_url and the read-only boundary (#25)

read_url returns text, and only text. This is a decision, now stated rather than implied: the fetched bytes are discarded after extraction, because saving them would make a reader tool write to the file system, and every tool of this server is read-only (readOnlyHint: true — all 18 of them).

To run search_text, inspect_structure, extract_tables, render_page or anything else against a PDF that lives at a URL, download the file first — with whatever fetch capability the calling environment has — and pass the local path. Fetching is the caller's responsibility, deliberately: an agent environment always has a way to download a file, and a reader that also writes files has stopped being a pure observer.

read_url remains the right tool for the one-shot question: what does the document at this URL say?

Pages can be rendered when text cannot be read (#23)

summarize reporting hasText: false used to be a dead end: nothing in this server could read the document any further. render_page closes that — it rasterises pages to PNG or JPEG and returns them as MCP image content blocks, so a vision model can read a scan, a diagram, a filled form, or handwriting.

render_page({ file_path: "/path/to/scan.pdf", pages: "1-3", format: "jpeg" })

pages is required: rendering is the most expensive operation here, and "all pages" of a 500-page scan should be a decision, not a default. The same 4 MB response budget as read_images applies, with omissions named.

Rendering runs on PDFium compiled to WebAssembly (@hyzyla/pdfium, an optional dependency). A WASM binary is the same bytes on every platform, so the published package still behaves identically wherever npx runs it — the reason native addons are not used here. Without the dependency installed, render_page reports what to install and every other tool works normally. PDFium (BSD-3-Clause) is a different engine from the pdf.js this server reads text with; the tool description says so, because a rendering difference between engines must not be attributed to the file.

Measured before choosing this: pdf.js + @napi-rs/canvas (1.0.7 and 0.1.80) segfaults the whole process on pages that draw images — exactly the pages this tool exists for — and renders blank pages when standardFontDataUrl is not configured.

Images come back as image files (#22)

read_images used to base64 imgData.data — pdfjs's decoded pixels. An 8×8 RGB image was 192 bytes with no PNG or JPEG signature anywhere in it, so the result could not be opened by any viewer and could not be read by a vision model, which is the reason to extract an image in the first place.

Images are now encoded (PNG by default, lossless; format: "jpeg" with quality when smaller matters) and returned as MCP image content blocks, with the metadata alongside in a text block. Both encoders are written out here — no native addon, no per-platform binary.

The response is bounded at 4 MB of encoded image data. A 200 dpi A4 scan is ~11.6 MB of pixels on its own, so images past the budget are named with the reason rather than dropped:

read_images({ file_path: "/path/to/scan.pdf", pages: "1", max_width: 1200, format: "jpeg" })

read_images returns the image XObjects a page draws. It is not a picture of the page — vector drawings and text are not covered by it.

Text extractability — three states, not two (#21)

read_text used to answer with text or with nothing, and nothing meant three different things. ISO 32000-2 §9.10.1 separates them, so this server does too. Every text-returning tool — read_text, read_url, search_text, extract_structured_text, summarize — reports, per page:

not_extractable is reported per font, so a page that mixes a readable font with an unreadable one is a partial loss and says so, rather than passing as complete.

The observation is made from the file, not from pdf.js's output: pdf.js synthesises a toUnicode map for every font it loads, so asking it whether a font has a /ToUnicode CMap answers yes for fonts whose dictionary has none.

Installation

npx (recommended)

npx @shuji-bonji/pdf-reader-mcp@latest

Claude Desktop

Add to your claude_desktop_config.json:

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

Use @latest. The -y flag in npx -y <pkg> only skips the install prompt — it does not check for updates. Without @latest, npx keeps running whichever version it cached the first time, so new releases never reach you. If you suspect you are on a stale version, run rm -rf ~/.npm/_npx and restart your client.

Claude Code

claude mcp add pdf-reader-mcp -- npx -y @shuji-bonji/pdf-reader-mcp@latest

From Source

git clone https://github.com/shuji-bonji/pdf-reader-mcp.gitcd pdf-reader-mcpnpm installnpm run build

Usage Examples

Get Page Count

get_page_count({ file_path: "/path/to/document.pdf" })→ 42
get_page_count({ file_path: "/path/to/document.pdf", response_format: "json" })→ { "pageCount": 42,    "scope": { "declaredCount": 42, "pagesReached": 42, "pageTree": { "status": "read" } } }

declaredCount is the /Count the page tree declares (ISO 32000-2 §7.7.3.2) and pagesReached is how many pages walking the tree actually reached. They differ only when the writer got /Count wrong; both are reported and neither is judged.

Metadata from both sources

get_metadata reads the eight document-level fields (title, author, subject, keywords, creator, producer, creation and modification dates) from the document information dictionary (§14.3.3) and the XMP metadata stream (§14.3.2). When both hold a value, PDF 2.0 and later use XMP (§14.3.3 deprecates the dictionary for this in PDF 2.0) and earlier versions use the dictionary. The two dates are the exception: they use the dictionary in every version, because §14.3.3 does not deprecate CreationDate and ModDate, and §14.5 compares PieceInfo's LastModified with the dictionary's ModDate. metadataSources names the source of each field, and metadataConflicts lists the fields whose two values differ (dates compared as instants). Values are returned as written — a dictionary date looks like D:20260915100000+09'00', an XMP date like 2026-09-15T10:00:00+09:00.

Search Text

search_text({  file_path: "/path/to/spec.pdf",  query: "digital signature",  pages: "1-20",  max_results: 10})→ Found 5 matches (page 3, 7, 12, 15, 18)

Summarize

summarize({ file_path: "/path/to/document.pdf" })→ | Pages | 42 |  | PDF Version | 2.0 |  | Tagged | Yes |  | Signatures | No |  | Images | 15 |

Compare Structure

compare_structure({  file_path_1: "/path/to/v1.pdf",  file_path_2: "/path/to/v2.pdf"})→ | Page Count  | 10 | 12 | ❌ |  | PDF Version | 1.7 | 2.0 | ❌ |  | Tagged      | true | true | ✅ |

Extract Structured Text (Tagged PDF, logical order)

extract_structured_text({ file_path: "/path/to/report.pdf", pages: "1-2" })→ # Structured Text  - **Tagged**: Yes / **Language**: en-US / **Elements**: 7
  ## Logical Content Order
  - **Document** (pages 1–2)    - **H1** (page 1) — Quarterly Report    - **P** (pages 1–2) — This paragraph begins on page one and continues on page two.    - **Table** (page 1)      | Item | Amount |      |---|---|      | Sales | 100 |    - **Figure** (page 2) — *alt:* A bar chart of sales

This answers "what is the text of the H1?" — which read_text (flat, coordinate order) cannot. Order is a depth-first traversal of the structure tree (ISO 32000-2 §14.8.2.5). /ActualText replaces the glyphs (§14.9.4), /Alt stays out of the body text (§14.9.3), list labels are reported separately, a list nested in an item comes back as its own L / LI elements after the item (not folded into the item's text), and an element spanning a page break stays ONE element. Use roles: ["H1", "H2"] to pull an outline. Untagged PDFs return isTagged: false with a reason — nothing is guessed from coordinates.

include_bbox: where the element is drawn

extract_structured_text({ file_path: "/doc.pdf", roles: ["P"], include_bbox: true })→ - **P** (page 1) — Measured paragraph    - *bbox* p1 `(50.0, 297.5, 161.4, 308.6)` — text-extent  - **Figure** (page 1)    - *bbox* p1 `(50.0, 150.0, 110.0, 190.0)` — layout-attribute-bbox

Rectangles are in PDF default user space (origin bottom-left, pt, normalised) — exactly what pdf-writer-mcp add_annotation takes, so "annotate this paragraph" needs no coordinate conversion in between. /Rotate and a shifted /CropBox do not move them.

basis says how strong the claim is, and the two are different in kind:

An element spanning pages gets ONE RECTANGLE PER PAGE — merging them would put content on a page it is not on. An element with no rectangle says why in boxNote rather than returning a zero-sized one.

A declaration is reported as-is, and cross-checked. Files state nonsense: the cover Figure of both Well-Tagged PDF 1.0 and the Tagged PDF Best Practice Guide declares /BBox [-32768 -32768 32767 32767] — int16 sentinels where a rectangle should be — and PDF32000_2008 has 131 of its 545 declarations reaching past the page edge. Since this output is meant to go straight into add_annotation, a declaration is checked against the page box (§7.7.3.3) and against the element's own text; either contradiction is reported in boxNote, with the rectangle still returned unaltered.

Measured against independent ground truth: on Well-Tagged PDF (WTPDF) 1.0, the 166 Link structure elements were compared with the 173 Link annotation /Rect values the producer placed for the same links — median IoU 0.972, none disjoint.

Extract Tables (Tagged PDF)

extract_tables({ file_path: "/path/to/kaisei-tsutatsu.pdf", pages: "1" })→ # Extracted Tables  - **Tagged**: Yes / **Pages Scanned**: 1 / **Tables Found**: 1
  ## Table 1 — Page 1
  | 改正後 | 改正前 |  | --- | --- |  | …第２条第 16 項《定義》… | …第２条第 15 項《定義》… |

A table that continues across a page break is reported as ONE table — pages is an array (e.g. ## Table 3 — Pages 5–7), and a table touching the requested pages range is returned whole. Cell text honours /ActualText replacements (as do read_text and search_text since #18). Untagged PDFs return an empty result with a note recommending the column-aware fallback below.

Read Untagged Multi-Column PDF

read_text({ file_path: "/path/to/older-shinkyu.pdf", split_columns: 2 })→ // Plain Y-sort would interleave columns://   "改正後セル1   改正前セル1\n 改正後セル2   改正前セル2..."//// With split_columns: 2 the left column is emitted first, then the right://   "改正後セル1\n改正後セル2\n…\n\n改正前セル1\n改正前セル2\n…"

Use split_columns: 2 | 3 for untagged multi-column PDFs. For Tagged PDFs with proper <Table> markup, extract_tables (above) is preferred.

Compact Whitespace (Japanese Forms)

read_text({ file_path: "/path/to/form.pdf", compact_whitespace: true })→ // Original PDF uses U+3000 fullwidth space as visual indentation://   " (   )   自   年   月   日   法   有 （   年   月   日）   有   有"//// With compact_whitespace: true://   "( ) 自 年 月 日 法 有 （ 年 月 日） 有 有"//// Empirically reduces character count by ~40% on form PDFs.

compact_whitespace is orthogonal to split_columns — both can be combined.

Tech Stack

• TypeScript + MCP TypeScript SDK
• pdfjs-dist (Mozilla) — text/image extraction, tag tree, annotations
• normativepdf + @normativepdf/recover — COS object access, read from files whose cross-reference table does not follow ISO 32000-2 §7.5
• Vitest — unit + E2E testing (483 tests)
• Biome — linting + formatting
• Zod — input validation

Testing

npm test              # Run all tests (unit + E2E: 483 tests)npm run test:e2e      # E2E tests only (283 tests)npm run test:watch    # Watch mode

Architecture

pdf-reader-mcp/├── src/│   ├── index.ts              # MCP Server entry point│   ├── constants.ts          # Shared constants│   ├── types.ts              # Type definitions│   ├── tools/│   │   ├── tier1/            # Basic tools (7)│   │   ├── tier2/            # Structure inspection (6)│   │   ├── tier3/            # Validation & analysis (3)│   │   └── index.ts          # Tool registration│   ├── services/│   │   ├── pdfjs-service.ts        # pdfjs-dist wrapper (parallel page processing)│   │   ├── recover-service.ts      # @normativepdf/recover — opening a document,│   │   │                           #   and how far it could be read (DocumentScope)│   │   ├── structure-service.ts    # Catalog, page tree, object statistics│   │   ├── font-service.ts         # Fonts named in each page's /Resources│   │   ├── signature-service.ts    # Signature fields of the AcroForm (structure only)│   │   ├── content-stream-service.ts  # Marked-content operators, text-showing tally│   │   ├── struct-tree-service.ts  # Logical structure (tags, structured text)│   │   ├── validation-service.ts   # Validation & comparison logic│   │   └── url-fetcher.ts          # URL fetching│   ├── schemas/              # Zod validation schemas│   └── utils/│       ├── pdf-helpers.ts    # PDF utilities (page range parsing, file I/O)│       ├── batch-processor.ts # Batch processing for large PDFs│       ├── formatter.ts      # Output formatting│       └── error-handler.ts  # Error handling└── tests/    ├── tier1/                # Unit tests    └── e2e/                  # E2E tests (16 suites, 283 tests)

Error Contract (houki-hub family)

Since v0.6.0, this MCP returns structured errors that follow the houki-hub family error contract, sharing a unified code vocabulary across the family. Combined with houki-egov-mcp / houki-nta-mcp, an LLM or Skill layer can interpret errors with consistent logic.

• docs/ERROR-CODES.md — error code vocabulary (houki-research-skill)
• docs/ERROR-HANDLING.md — handling policy / next_actions templates

Implementation is independent — no dependency on houki-abbreviations or other family packages. The reference implementation is houki-egov-mcp/src/errors.ts; pdf-reader-mcp's local definition is in src/errors.ts.

On error, every tool returns isError: true and the JSON-stringified LawServiceError in content[0].text:

{  "error": "The file does not appear to be a valid PDF.",  "code": "INVALID_PDF",  "retryable": false,  "hint": "The file breaks a structural rule of ISO 32000 (for example, it does not begin with %PDF- (§7.5.2), or it does not end with startxref and %%EOF (§7.5.5)), so it cannot be read as written. This is a finding about the file, not a failure of this server.",  "next_actions": [    {      "action": "inspect_structure",      "reason": "Look at the catalog and the page tree to see where the file stops following the specification."    }  ],  "detail": { "cause": "The PDF file is empty, i.e. its size is zero bytes." }}

Every error carries error (an English sentence), code and retryable; hint, next_actions and detail are optional, and nothing else appears at the top level (PDF family contract, pdf-agent-stack docs/specs/06 §2.3, since v0.19.0). retryable is fixed per code: it says whether changing the arguments, flags or environment could make the same reading of the same document succeed. The values of next_actions[].action are stable — a Skill may branch on them.

Codes used by pdf-reader-mcp

Migration note (v0.5.x → v0.6.0)

旧 v0.5.x までは content[0].text に Error: ...\n\nSuggestion: ... という人間可読文字列を入れていました。v0.6.0 では同じ場所に JSON 文字列 が入ります。LLM 側でテキスト解釈に依存していた場合は、JSON.parse(content[0].text) での解釈に切り替えてください。isError: true フラグで構造化エラーかどうかを判定できます。

Pairing with pdf-spec-mcp

pdf-spec-mcp provides PDF specification knowledge (ISO 32000-2, etc.). With both servers enabled, an LLM can perform specification-aware workflows:

1. summarize — get a PDF overview
2. inspect_tags — examine the tag structure
3. pdf-spec-mcp get_requirements — fetch PDF/UA requirements
4. pdf-verify-mcp validate_conformance — check conformance (flavour: "pdfua-1")
5. compare_structure — diff before/after fixes

License

MIT