PDF Reader MCP 📄
Production-ready PDF processing server for AI agents
5-10x faster parallel processing • Y-coordinate content ordering • 94%+ test coverage • 103 tests passing
基于原项目: 此项目基于 pdf-reader-mcp 修改而来
🚀 Overview
PDF Reader MCP is a production-ready Model Context Protocol server that empowers AI agents with enterprise-grade PDF processing capabilities. Extract text, images, and metadata with unmatched performance and reliability.
The Problem:
// Traditional PDF processing
- Sequential page processing (slow)
- No natural content ordering
- Complex path handling
- Poor error isolation
The Solution:
// PDF Reader MCP
- 5-10x faster parallel processing ⚡
- Y-coordinate based ordering 📐
- Flexible path support (absolute/relative) 🎯
- Per-page error resilience 🛡️
- 94%+ test coverage ✅
Result: Production-ready PDF processing that scales.
⚡ Key Features
Performance
- 🚀 5-10x faster than sequential with automatic parallelization
- ⚡ 12,933 ops/sec error handling, 5,575 ops/sec text extraction
- 💨 Process 50-page PDFs in seconds with multi-core utilization
- 📦 Lightweight with minimal dependencies
Developer Experience
- 🎯 Path Flexibility - Absolute & relative paths, Windows/Unix support (v1.3.0)
- 🖼️ Smart Ordering - Y-coordinate based content preserves document layout
- 🛡️ Type Safe - Full TypeScript with strict mode enabled
- 📚 Battle-tested - 103 tests, 94%+ coverage, 98%+ function coverage
- 🎨 Simple API - Single tool handles all operations elegantly
📊 Performance Benchmarks
Real-world performance from production testing:
| Operation | Ops/sec | Performance | Use Case |
|---|---|---|---|
| Error handling | 12,933 | ⚡⚡⚡⚡⚡ | Validation & safety |
| Extract full text | 5,575 | ⚡⚡⚡⚡ | Document analysis |
| Extract page | 5,329 | ⚡⚡⚡⚡ | Single page ops |
| Multiple pages | 5,242 | ⚡⚡⚡⚡ | Batch processing |
| Metadata only | 4,912 | ⚡⚡⚡ | Quick inspection |
Parallel Processing Speedup
| Document | Sequential | Parallel | Speedup |
|---|---|---|---|
| 10-page PDF | ~2s | ~0.3s | 5-8x faster |
| 50-page PDF | ~10s | ~1s | 10x faster |
| 100+ pages | ~20s | ~2s | Linear scaling with CPU cores |
Benchmarks vary based on PDF complexity and system resources.
📦 Installation
# Quick start - zero installation
npx @sylphx/pdf-reader-mcp
# Using pnpm (recommended)
pnpm add @sylphx/pdf-reader-mcp
# Using npm
npm install @sylphx/pdf-reader-mcp
# Using yarn
yarn add @sylphx/pdf-reader-mcp
# For Claude Desktop (easiest)
npx -y @smithery/cli install @sylphx/pdf-reader-mcp --client claude
🎯 Quick Start
Configuration
Add to your MCP client (claude_desktop_config.json, Cursor, Cline):
{
"mcpServers": {
"pdf-reader-mcp": {
"command": "npx",
"args": ["@bachstudio/pdf-reader-mcp"]
}
}
}
Basic Usage
{
"sources": [{
"path": "documents/report.pdf"
}],
"include_full_text": true,
"include_metadata": true,
"include_page_count": true
}
Result:
- ✅ Full text content extracted
- ✅ PDF metadata (author, title, dates)
- ✅ Total page count
- ✅ Structural sharing - unchanged parts preserved
Extract Specific Pages
{
"sources": [{
"path": "documents/manual.pdf",
"pages": "1-5,10,15-20"
}],
"include_full_text": true
}
Absolute Paths (v1.3.0+)
// Windows - Both formats work!
{
"sources": [{
"path": "C:\\Users\\John\\Documents\\report.pdf"
}],
"include_full_text": true
}
// Unix/Mac
{
"sources": [{
"path": "/home/user/documents/contract.pdf"
}],
"include_full_text": true
}
No more "Absolute paths are not allowed" errors!
Extract Images with Natural Ordering
{
"sources": [{
"path": "presentation.pdf",
"pages": [1, 2, 3]
}],
"include_images": true,
"include_full_text": true
}
Response includes:
- Text and images in exact document order (Y-coordinate sorted)
- Base64-encoded images with metadata (width, height, format)
- Natural reading flow preserved for AI comprehension
Batch Processing
{
"sources": [
{ "path": "C:\\Reports\\Q1.pdf", "pages": "1-10" },
{ "path": "/home/user/Q2.pdf", "pages": "1-10" },
{ "url": "https://example.com/Q3.pdf" }
],
"include_full_text": true
}
⚡ All PDFs processed in parallel automatically!
✨ Features
Core Capabilities
- ✅ Text Extraction - Full document or specific pages with intelligent parsing
- ✅ Image Extraction - Base64-encoded with complete metadata (width, height, format)
- ✅ Content Ordering - Y-coordinate based layout preservation for natural reading flow
- ✅ Metadata Extraction - Author, title, creation date, and custom properties
- ✅ Page Counting - Fast enumeration without loading full content
- ✅ Dual Sources - Local files (absolute or relative paths) and HTTP/HTTPS URLs
- ✅ Batch Processing - Multiple PDFs processed concurrently
Advanced Features
- ⚡ 5-10x Performance - Parallel page processing with Promise.all
- 🎯 Smart Pagination - Extract ranges like "1-5,10-15,20"
- 🖼️ Multi-Format Images - RGB, RGBA, Grayscale with automatic detection
- 🛡️ Path Flexibility - Windows, Unix, and relative paths all supported (v1.3.0)
- 🔍 Error Resilience - Per-page error isolation with detailed messages
- 📏 Large File Support - Efficient streaming and memory management
- 📝 Type Safe - Full TypeScript with strict mode enabled
🆕 What's New in v1.3.0
🎉 Absolute Paths Now Supported!
// ✅ Windows
{ "path": "C:\\Users\\John\\Documents\\report.pdf" }
{ "path": "C:/Users/John/Documents/report.pdf" }
// ✅ Unix/Mac
{ "path": "/home/john/documents/report.pdf" }
{ "path": "/Users/john/Documents/report.pdf" }
// ✅ Relative (still works)
{ "path": "documents/report.pdf" }
Other Improvements:
- 🐛 Fixed Zod validation error handling
- 📦 Updated all dependencies to latest versions
- ✅ 103 tests passing, 94%+ coverage maintained
📋 View Full Changelog
v1.2.0 - Content Ordering
- Y-coordinate based text and image ordering
- Natural reading flow for AI models
- Intelligent line grouping
v1.1.0 - Image Extraction & Performance
- Base64-encoded image extraction
- 10x speedup with parallel processing
- Comprehensive test coverage (94%+)
📖 API Reference
read_pdf Tool
The single tool that handles all PDF operations.
Parameters
| Parameter | Type | Description | Default |
|---|---|---|---|
sources | Array | List of PDF sources to process | Required |
include_full_text | boolean | Extract full text content | false |
include_metadata | boolean | Extract PDF metadata | true |
include_page_count | boolean | Include total page count | true |
include_images | boolean | Extract embedded images | false |
Source Object
{
path?: string; // Local file path (absolute or relative)
url?: string; // HTTP/HTTPS URL to PDF
pages?: string | number[]; // Pages to extract: "1-5,10" or [1,2,3]
}
Examples
Metadata only (fast):
{
"sources": [{ "path": "large.pdf" }],
"include_metadata": true,
"include_page_count": true,
"include_full_text": false
}
From URL:
{
"sources": [{
"url": "https://arxiv.org/pdf/2301.00001.pdf"
}],
"include_full_text": true
}
Page ranges:
{
"sources": [{
"path": "manual.pdf",
"pages": "1-5,10-15,20" // Pages 1,2,3,4,5,10,11,12,13,14,15,20
}]
}
🔧 Advanced Usage
📐 Y-Coordinate Content Ordering
Content is returned in natural reading order based on Y-coordinates:
Document Layout:
┌─────────────────────┐
│ [Title] Y:100 │
│ [Image] Y:150 │
│ [Text] Y:400 │
│ [Photo A] Y:500 │
│ [Photo B] Y:550 │
└─────────────────────┘
Response Order:
[
{ type: "text", text: "Title..." },
{ type: "image", data: "..." },
{ type: "text", text: "..." },
{ type: "image", data: "..." },
{ type: "image", data: "..." }
]
Benefits:
- AI understands spatial relationships
- Natural document comprehension
- Perfect for vision-enabled models
- Automatic multi-line text grouping
🖼️ Image Extraction
Enable extraction:
{
"sources": [{ "path": "manual.pdf" }],
"include_images": true
}
Response format:
{
"images": [{
"page": 1,
"index": 0,
"width": 1920,
"height": 1080,
"format": "rgb",
"data": "base64-encoded-png..."
}]
}
Supported formats: RGB, RGBA, Grayscale Auto-detected: JPEG, PNG, and other embedded formats
📂 Path Configuration
Absolute paths (v1.3.0+) - Direct file access:
{ "path": "C:\\Users\\John\\file.pdf" }
{ "path": "/home/user/file.pdf" }
Relative paths - Workspace files:
{ "path": "docs/report.pdf" }
{ "path": "./2024/Q1.pdf" }
Configure working directory:
{
"mcpServers": {
"pdf-reader-mcp": {
"command": "npx",
"args": ["@sylphx/pdf-reader-mcp"],
"cwd": "/path/to/documents"
}
}
}
📊 Large PDF Strategies
Strategy 1: Page ranges
{ "sources": [{ "path": "big.pdf", "pages": "1-20" }] }
Strategy 2: Progressive loading
// Step 1: Get page count
{ "sources": [{ "path": "big.pdf" }], "include_full_text": false }
// Step 2: Extract sections
{ "sources": [{ "path": "big.pdf", "pages": "50-75" }] }
Strategy 3: Parallel batching
{
"sources": [
{ "path": "big.pdf", "pages": "1-50" },
{ "path": "big.pdf", "pages": "51-100" }
]
}
🔧 Troubleshooting
"Absolute paths are not allowed"
Solution: Upgrade to v1.3.0+
npm update @sylphx/pdf-reader-mcp
Restart your MCP client completely.
"File not found"
Causes:
- File doesn't exist at path
- Wrong working directory
- Permission issues
Solutions:
Use absolute path:
{ "path": "C:\\Full\\Path\\file.pdf" }
Or configure cwd:
{
"pdf-reader-mcp": {
"command": "npx",
"args": ["@sylphx/pdf-reader-mcp"],
"cwd": "/path/to/docs"
}
}
"No tools showing up"
Solution:
npm cache clean --force
rm -rf node_modules package-lock.json
npm install @sylphx/pdf-reader-mcp@latest
Restart MCP client completely.
🏗️ Architecture
Tech Stack
| Component | Technology |
|---|---|
| Runtime | Node.js 22+ ESM |
| PDF Engine | PDF.js (Mozilla) |
| Validation | Zod + JSON Schema |
| Protocol | MCP SDK |
| Language | TypeScript (strict) |
| Testing | Vitest (103 tests) |
| Quality | Biome (50x faster) |
| CI/CD | GitHub Actions |
Design Principles
- 🔒 Security First - Flexible paths with secure defaults
- 🎯 Simple Interface - One tool, all operations
- ⚡ Performance - Parallel processing, efficient memory
- 🛡️ Reliability - Per-page isolation, detailed errors
- 🧪 Quality - 94%+ coverage, strict TypeScript
- 📝 Type Safety - No
anytypes, strict mode - 🔄 Backward Compatible - Smooth upgrades always
🧪 Development
Setup & Scripts
Prerequisites:
- Node.js >= 22.0.0
- pnpm (recommended) or npm
Setup:
git clone https://github.com/SylphxAI/pdf-reader-mcp.git
cd pdf-reader-mcp
pnpm install && pnpm build
Scripts:
pnpm run build # Build TypeScript
pnpm run test # Run 103 tests
pnpm run test:cov # Coverage (94%+)
pnpm run check # Lint + format
pnpm run check:fix # Auto-fix
pnpm run benchmark # Performance tests
Quality:
- ✅ 103 tests
- ✅ 94%+ coverage
- ✅ 98%+ function coverage
- ✅ Zero lint errors
- ✅ Strict TypeScript
Contributing
Quick Start:
- Fork repository
- Create branch:
git checkout -b feature/awesome - Make changes:
pnpm test - Format:
pnpm run check:fix - Commit: Use Conventional Commits
- Open PR
Commit Format:
feat(images): add WebP support
fix(paths): handle UNC paths
docs(readme): update examples
See CONTRIBUTING.md
📚 Documentation
- 📖 Full Docs - Complete guides
- 🚀 Getting Started - Quick start
- 📘 API Reference - Detailed API
- 🏗️ Design - Architecture
- ⚡ Performance - Benchmarks
- 🔍 Comparison - vs. alternatives
🗺️ Roadmap
✅ Completed
- Image extraction (v1.1.0)
- 5-10x parallel speedup (v1.1.0)
- Y-coordinate ordering (v1.2.0)
- Absolute paths (v1.3.0)
- 94%+ test coverage (v1.3.0)
🚀 Next
- OCR for scanned PDFs
- Annotation extraction
- Form field extraction
- Table detection
- 100+ MB streaming
- Advanced caching
- PDF generation
Vote at Discussions
🏆 Recognition
Featured on:
Trusted worldwide • Enterprise adoption • Battle-tested
🤝 Support
- 🐛 Bug Reports
- 💬 Discussions
- 📖 Documentation
Show Your Support: ⭐ Star • 👀 Watch • 🐛 Report bugs • 💡 Suggest features • 🔀 Contribute
📊 Stats
103 Tests • 94%+ Coverage • Production Ready
📄 License
MIT © Sylphx
🙏 Credits
Built with:
Special thanks to the open source community ❤️
5-10x faster. Production-ready. Battle-tested.
The PDF processing server that actually scales
sylphx.com •
@SylphxAI •
hi@sylphx.com