PDF MCP Server

An MCP server that enables reading PDF file contents, allowing PDF documents to be used as a knowledge base for LLMs.

Features

• High-Quality Extraction: Uses marker-pdf (via a Python backend) to extract text with layout awareness and high-fidelity LaTeX equation recognition.

• Robust Fallback: Automatically switches to a Node.js-based parser (pdf-parse) if the Python environment is unavailable or fails, ensuring extraction always succeeds (albeit with lower formatting quality).

• Smart Filtering: Supports page range extraction to process only relevant sections of large documents.

Installation

Prerequisites

• Node.js (v18+)

• Python (v3.10+) and pip (for high-quality extraction)

Setup

1. Install Node.js dependencies:

   npm install

2. Install Python dependencies (Recommended): To enable high-quality extraction (especially for scientific papers with math), install the Python dependencies.

   # Create or activate a virtual environment if desired python3 -m pip install -r python/requirements.txt

   Note: The first time you run the tool with the Python backend, it will download necessary AI models (OCR, layout analysis, etc.) to a local cache. This download is approximately 3.3GB. Ensure you have a stable internet connection.

3. Build the server:

   npm run build

Usage

Configuration for Claude/MCP Clients

Add this to your MCP settings configuration:

Tool: read_pdf

Reads and extracts text content from a PDF file.

Inputs:

• path (string): Absolute path to the PDF file.

• start_page (number, optional): Starting page number (1-based).

• end_page (number, optional): Ending page number (1-based).

How it works:

1. Attempt 1 (Python/Marker): The server tries to run the internal convert.py script.

   • If successfully configured, this loads the marker models from the local cache (.cache directory in the project).

   • It accurately converts equations to LaTeX and preserves document structure.

2. Attempt 2 (Fallback): If the Python script fails (e.g., missing dependencies, runtime error), the server catches the error and uses pdf-parse (a native Node.js library).

   • This extracts raw text. Equations may appear as linearized text, and layout may be less preserved.

Troubleshooting

• Permission Errors: The project is configured to use a local .cache directory for models to avoid system permission issues. If you encounter errors, ensure the project directory is writable.

• Slow Performance: The high-quality extraction uses deep learning models. It can be slow on large documents without a GPU. Use the start_page and end_page arguments to extract only what you need.