Paper-render

Automated pipeline: PDF paper → beautiful HTML reading notes. Powered by Claude Code + MCP Server.

Features

• PDF Full-text Parsing: Text extraction, formula recognition (Nougat OCR), figure extraction
• Markdown Notes Generation: Structured Chinese reading notes with formulas, figure references, and critical analysis
• HTML Presentation Generation: Single-file HTML with dark/light theme toggle, bilingual (CN/EN), KaTeX formula rendering, responsive sidebar navigation

Project Structure

Paper-render/
├── .claude/
│   └── commands/
│       └── paper-read.md        # Claude Code skill (one-click paper notes)
├── pdf-tools-mcp/               # PDF analysis MCP Server
│   ├── pyproject.toml
│   └── src/pdf_tools_mcp/
│       └── server.py
├── templates/                   # HTML template resources
│   ├── base.css                 # Shared CSS (dark/light themes, components)
│   ├── base.js                  # Shared JS (theme toggle, language toggle, TOC, image replacement)
│   └── skeleton.html            # HTML skeleton template (available components & structure)
├── Paper-Library/               # Generated paper notes
│   ├── LPFM/
│   │   ├── LPFM.pdf
│   │   ├── LPFM_notes.md
│   │   ├── LPFM_presentation.html
│   │   └── figures/
│   └── PixCell/
│       ├── PixCell.pdf
│       ├── PixCell_notes.md
│       ├── PixCell_presentation.html
│       └── figures/
└── README.md

Installation

1. Install pdf-tools MCP Server

cd pdf-tools-mcp

# Basic install (text extraction, page rendering, image extraction)
pip install -e .

# Full install (with Nougat formula recognition, requires GPU)
pip install -e ".[nougat]"

2. Register MCP Server with Claude Code

# Option 1: Direct registration
claude mcp add pdf-tools -- pdf-tools-mcp

# Option 2: Using uvx (no install needed)
claude mcp add pdf-tools -- uvx --from /path/to/pdf-tools-mcp pdf-tools-mcp

Verify registration:

# List registered MCP servers
claude mcp list

3. Confirm skill availability

The skill file is located at .claude/commands/paper-read.md. Claude Code automatically detects commands in the project directory.

Type / in Claude Code to see the paper-read command.

Usage

One-click Generation (Recommended)

In Claude Code, run:

/paper-read /path/to/your/paper.pdf

Claude will automatically:

1. Read the full PDF text (batch text extraction, formula recognition)
2. Extract all important figures from the paper
3. Generate structured Markdown reading notes
4. Generate a beautiful bilingual HTML presentation page

All outputs are saved in Paper-Library/{paper-name}/.

Manual MCP Tool Usage

You can also call MCP tools directly in Claude Code:

# View PDF info
Use pdf_info to check basic info of paper.pdf

# Extract text
Use pdf_read_text to read pages 1-10 of paper.pdf

# Formula recognition (requires nougat + GPU)
Use pdf_read_formulas to read formulas on page 3 of paper.pdf

# One-step auto-extract all figures (recommended)
Use pdf_extract_figures to extract all figures from paper.pdf to figures/ directory

# Detect figure regions (metadata only, no rendering)
Use pdf_detect_figures to detect figure positions on pages 3-8 of paper.pdf

# Manually crop a specific region
Use pdf_render_region to render region (30,60)-(565,530) on page 3 of paper.pdf

# Render full page as image
Use pdf_render_page to render page 5 of paper.pdf

# Extract embedded images (raw image layers)
Use pdf_extract_images to extract embedded images from paper.pdf

MCP Server Tools

Tool	Description	Dependency
pdf_info	PDF metadata (page count, title, author, TOC)	Basic
pdf_read_text	Fast text extraction with table support	Basic
pdf_read_formulas	Formula/LaTeX recognition (Nougat OCR)	[nougat] + GPU
pdf_extract_figures	Smart figure extraction: auto-detect + cluster + crop (recommended)	Basic
pdf_detect_figures	Detect figure regions, return metadata (no rendering)	Basic
pdf_render_region	Render a specific rectangular region of a page (manual fine-tuning)	Basic
pdf_render_page	Render full page as PNG image	Basic
pdf_extract_images	Extract raw embedded image layers from PDF	Basic

Figure Extraction Workflow

Recommended workflow for figure extraction:

1. pdf_extract_figures (preferred) — One-step auto-detect and crop all figures
   • Supports both raster images and pure vector graphics (flowcharts, diagrams)
   • Auto-clusters sub-panels belonging to the same figure
   • Crop regions include axes, labels, and vector annotations
2. pdf_detect_figures — Detect only (no rendering), for preview and debugging
3. pdf_render_region — Manually specify a rectangular region, for fine-tuning inaccurate auto-crops

HTML Presentation Features

• Dark/Light Theme: One-click toggle, preference saved to localStorage
• Bilingual (CN/EN): All content in both Chinese and English, one-click language switch
• KaTeX Formulas: Inline $...$ and display $$...$$ math formulas
• Responsive Sidebar Navigation: Fixed left TOC on desktop, collapsible menu on mobile
• Scroll Highlighting: Current reading position auto-highlighted in TOC
• Rich Components: Cards, hint boxes, flowcharts, data tables, bar charts, metric highlights, collapsible discussions, etc.
• Lazy Image Loading: Placeholders auto-replaced with extracted paper figures
• Single-file Deployment: CSS/JS fully inlined, opens directly in browser

Customizing Templates

Template files are in templates/:

• base.css — Modify theme colors, component styles
• base.js — Modify interaction behavior
• skeleton.html — View available HTML components and structural patterns

Dependencies

• Python >= 3.10
• PyMuPDF >= 1.24.0
• MCP >= 1.0.0
• Pillow
• (Optional) Nougat OCR — Formula recognition
• (Optional) CUDA GPU — Accelerate Nougat inference