Extract figures and tables from PDF documents using this FastAPI-based service. The Figure Extractor API and MCP Server provides a straightforward HTTP interface for PDFFigures 2.0, a robust figure extraction system developed by the Allen Institute for AI.
This API wrapper makes it ideal for integration into various applications and workflows, particularly for Retrieval-Augmented Generation (RAG) applications.
The MCP Server, powered by FastMCP, exposes the PDF extraction functionality as a MCP tool. This allows for seamless integration with AI Agents and workflows that can automatically call the extraction service.
- The server hosts the
extract_figures_from_pdftool, which can be invoked via an HTTP request to the/mcpendpoint. This tool takes a PDF URL, processes the document, and returns the extracted figures and tables in a structured JSON format.
The default MCP Server url:
http://localhost:5001/mcp
This API service is built on top of PDFFigures 2.0, a Scala-based project by the Allen Institute for AI. PDFFigures 2.0 is specifically designed to extract figures, captions, tables, and section titles from scholarly documents in computer science domain. The original work is described in their academic paper: "PDFFigures 2.0: Mining Figures from Research Papers" (Clark and Divvala, 2016). You can read the paper here and visit the PDFFigures 2.0 website.
┌─────────────────┐ ┌──────────────────┐ ┌────────────────┐
│ Your App │ HTTP │ Figure Extractor │ JNI │ PDFFigures │
│ (Any Language) │──────► API & MCP Server│──────► 2.0 │
│ │ │ Python(FastAPI) │ │ (Scala/JVM) │
└─────────────────┘ └──────────────────┘ └────────────────┘
- PDF figure and table extraction
- Support for local and remote PDF files
- Statistics of the extracted tables and figures
- Docker support for easy deployment
- Visualization options for PDF parsing
This project now includes a FastMCP tool that allows calling the PDF extraction service programmatically. The extract_figures_from_pdf tool can be used to extract figures from a PDF given a URL.
-
Machine Learning Dataset Creation Extract visual data from clinical trial reports and research papers to build training datasets for medical image analysis and AI models, enabling researchers to efficiently aggregate figures for training machine learning algorithms in healthcare diagnostics.
-
Clinical Research Data Mining Automatically extract and catalog figures from medical research articles, capturing key visualizations like treatment effect graphs, patient outcome charts, and experimental result diagrams to support systematic reviews and meta-analysis.
-
Academic Literature Review and Education Quickly compile comprehensive visual libraries from academic publications, allowing researchers and educators to create teaching resources, compare research methodologies, and track visual trends across scientific disciplines.
Build and start the extraction server:
docker build -t pdffigures2 .
docker run -d -p 5001:5001 pdffigures2The image is ~286MB — optimized via Alpine base, jlink minimal JRE, and pip cleanup:
| Metric | Original | Optimized | Change |
|---|---|---|---|
| Base Image | slim | alpine | -70MB |
| JRE | apt | jlink | +42MB |
| pip Packages | Original | Cleaned | -24MB |
| System Packages | apt | apk | -12MB |
| Image Size | 452MB | 286MB | -166MB (-37%) |
-
Open
http://localhost:5001/docsto view the API documentation. -
Use
curlto test extraction from a PDF URL:curl -X POST http://localhost:5001/api/extract \ -F "pdf_url=https://example.com/sample.pdf"
An API client that calls the extraction server and downloads rendered figures locally. Designed for agent environments without MCP support.
# Copy config and point to the server
cp .env.example .env
# (edit .env if the server is remote)
# Run extraction against a local PDF
skills/pdffigures2/scripts/pdffigures2 paper.pdf -o ./extracted/The script POSTs the PDF to PDFFIGURES2_API_URL, downloads each rendered figure, and prints a structured JSON summary to stdout. See skills/pdffigures2/SKILL.md for the full agent-facing documentation.
The /api/extract endpoint now supports both PDF file uploads and PDF URLs.
You can send a POST request to /api/extract with either:
- A
file(multipart/form-data) containing the PDF. - A
pdf_url(form-data) containing the URL of the PDF.
The API will return a JSON response with extracted figures and tables, including full renderURL paths.
A JSON response example (imageText truncated for brevity):
[
{
"caption": "TABLE III CMAPSS DATASET ATTRIBUTES",
"captionBoundary": {
"x1": 113.12599182128906,
"x2": 225.1184844970703,
"y1": 116.80506896972656,
"y2": 130.57305908203125
},
"figType": "Table",
"imageText": [
"Required", "fan", "conversion", "speed", "rpm", "High-pressure", "turbines", "cool", "air", "flow",
"lbm/s", "Low-pressure", "turbines", "cool", "air", "flow", "lbm/s", "Bleed", "enthalpy", "-", "Required"
],
"name": "III",
"page": 5,
"regionBoundary": {
"x1": 46.8,
"x2": 291.12,
"y1": 140.88,
"y2": 385.91999999999996
},
"renderDpi": 300,
"renderURL": "http://localhost:5001/resources/4-TableIII-1.png"
},
{
"caption": "Fig. 3. Agentic AI implementation with Google ADK.",
"captionBoundary": {
"x1": 335.4129943847656,
"x2": 516.9002685546875,
"y1": 228.6050567626953,
"y2": 233.40704345703125
},
"figType": "Figure",
"imageText": [],
"name": "3",
"page": 5,
"regionBoundary": {
"x1": 302.88,
"x2": 549.12,
"y1": 86.88,
"y2": 216
},
"renderDpi": 300,
"renderURL": "http://localhost:5001/resources/4-Figure3-1.png"
},
{
"caption": "TABLE I COMPARISON BETWEEN AI AGENTS AND AGENTIC AI",
"captionBoundary": {
"x1": 204.9189910888672,
"x2": 390.3569641113281,
"y1": 54.10902404785156,
"y2": 67.87701416015625
},
"figType": "Table",
"imageText": [
"pert", "systems", "LLM-based", "agents,", "multi-agent", "coordination,", "intent-based", "workflows", "Task",
"Scope", "Focused", "on", "short-term,", "well-defined", "tasks", "Oriented", "toward", "long-term,", "dynamic,",
],
"name": "I",
"page": 2,
"regionBoundary": {
"x1": 51.839999999999996,
"x2": 543.12,
"y1": 77.75999999999999,
"y2": 217.92
},
"renderDpi": 300,
"renderURL": "http://localhost:5001/resources/4-TableI-1.png"
},
{
"caption": "Fig. 1. Traditional AI Agent vs. Agentic AI",
"captionBoundary": {
"x1": 224.30499267578125,
"x2": 370.9708251953125,
"y1": 448.1660461425781,
"y2": 452.968017578125
},
"figType": "Figure",
"imageText": [],
"name": "1",
"page": 2,
"regionBoundary": {
"x1": 45.839999999999996,
"x2": 549.12,
"y1": 230.88,
"y2": 436.08
},
"renderDpi": 300,
"renderURL": "http://localhost:5001/resources/4-Figure1-1.png"
},
{
"caption": "TABLE IV SUMMARY OF ENGINE MAINTENANCE ACTIONS",
"captionBoundary": {
"x1": 215.99301147460938,
"x2": 379.2913513183594,
"y1": 54.10902404785156,
"y2": 67.87701416015625
},
"figType": "Table",
"imageText": [
"#", "Engines", "RUL", "Range", "Recommended", "Action", "Priority", "Cost", "(USD)", "Labor", "Hours", "Assigned",
"Staff", "Scheduled", "Time", "15", "82–124", "MONITOR", "low", "0", "0", "[jr", "mechanic]", "Within", "7", "days"
],
"name": "IV",
"page": 7,
"regionBoundary": {
"x1": 48.96,
"x2": 549.12,
"y1": 77.75999999999999,
"y2": 135.12
},
"renderDpi": 300,
"renderURL": "http://localhost:5001/resources/4-TableIV-1.png"
},
{
"caption": "TABLE II KEY COMPONENTS OF INTENTION PROCESSING",
"captionBoundary": {
"x1": 345.09295654296875,
"x2": 507.22100830078125,
"y1": 417.8482666015625,
"y2": 431.6162414550781
},
"figType": "Table",
"imageText": [
"Targets", "Specify", "the", "resources", "or", "entities", "to", "which", "the", "intent", "applies.", "Can", "be",
"defined", "statically", "(explicit", "list)", "or", "dynamically", "(using", "filters", "or", "criteria).", "Context"
],
"name": "II",
"page": 3,
"regionBoundary": {
"x1": 302.88,
"x2": 549.12,
"y1": 441.84,
"y2": 677.04
},
"renderDpi": 300,
"renderURL": "http://localhost:5001/resources/4-TableII-1.png"
},
{
"caption": "Fig. 2. Proposed framework for Industry 5.0 applying intent-based and Agentic AI.",
"captionBoundary": {
"x1": 159.51600646972656,
"x2": 435.7596435546875,
"y1": 288.3310241699219,
"y2": 293.13299560546875
},
"figType": "Figure",
"imageText": [],
"name": "2",
"page": 4,
"regionBoundary": {
"x1": 45.839999999999996,
"x2": 549.12,
"y1": 49.68,
"y2": 276
},
"renderDpi": 300,
"renderURL": "http://localhost:5001/resources/4-Figure2-1.png"
}
]Clone the repository if you want to modify the code or run locally without Docker:
git clone https://github.com/vlln/pdffigures-mcp-server.git
cd pdffigures-mcp-serverfigure_extractor.py is a CLI tool for testing extraction against a running API server. It sends a local PDF to the API and downloads the extracted figures:
# Start the server first (via Docker or uvicorn), then:
python figure_extractor.py <path-to-pdf> --output_dir ./outputOptions:
--output_dir— Directory to save downloaded figures (default:./output)--url— Custom API endpoint (default:http://localhost:5001/api/extract)
project/
├── Dockerfile # Multi-stage Alpine image (server + jlink JRE + pdffigures2 JAR)
├── .dockerignore
├── .env.example # Example config for the Agent Skill
├── app/ # FastAPI web service code
│ ├── __init__.py
│ ├── app.py # API endpoints and MCP Server
│ ├── service.py # Runs pdffigures2 JAR via subprocess
│ └── utils.py # File I/O helpers
├── skills/ # Agent Skills (non-MCP alternative)
│ └── pdffigures2/ # API client wrapper for pdffigures2
├── figure_extractor.py # CLI tool for testing extraction
└── README.md
This project is a fork of Huang-lab/figure-extractor. We are grateful to the original authors for their work.
This project is licensed under the Apache License 2.0.