PDF AI Extractor MCP

A dual-mode Apify Actor that downloads PDFs, extracts clean text, and analyzes the content using major AI models. It can run as a one-time job or as a persistent MCP (Model Context Protocol) server for AI agents.

Overview

This actor solves the common problem of extracting and analyzing unstructured data from PDFs. It reliably downloads a PDF, parses and cleans the text, and then sends that text to your chosen AI provider (OpenAI, Google Gemini, or Anthropic) for analysis based on your prompt. It operates in two distinct modes: a standard "normal" mode for single-run automation, and an "MCP" mode where it functions as a WebSocket server, allowing AI agents like those in ChatGPT or Claude to call its functions as tools.

Key Features

• Robust PDF Extraction: Downloads PDFs from a public URL and uses pdf-parse to extract and clean the text.
• Multi-AI Support: Connects to OpenAI (GPT-4, GPT-4o, o3-mini), Google Gemini (1.5 Flash/Pro), or Anthropic Claude (Haiku, Sonnet, Opus) APIs.
• Dual Operation Modes:
  • Normal Mode: Runs a single extraction and analysis job, returning structured JSON. Ideal for automated backend workflows.
  • MCP Mode: Launches a WebSocket server that exposes tools (e.g., extractPdf(), analyze()) to AI agents via the MCP standard.
• Structured Output: Returns the AI's analysis in a consistent JSON format.

How to Use

Prerequisites

Set the required environment variables for your chosen AI provider(s) in a .env file:

OPENAI_API_KEY=your_key_here
GEMINI_API_KEY=your_key_here
ANTHROPIC_API_KEY=your_key_here
MCP_PORT=8080 # Optional, for MCP mode

Running the Actor

Normal Mode (One-time job):
Provide an input object specifying "mode": "normal", the PDF URL, AI provider, and your analysis prompt.

apify run --purge --input-file=tests/input.normal.json

MCP Mode (Agent Server):
Provide an input object specifying "mode": "mcp". The actor will start a WebSocket server (default: ws://localhost:8080) that AI agents can connect to.

apify run --purge --input-file=tests/input.mcp.json

Input / Output

Input Schema (Normal Mode)

Example Input (Normal Mode):

{
  "mode": "normal",
  "pdfUrl": "https://example.com/document.pdf",
  "aiProvider": "openai",
  "prompt": "Summarize the key points from this contract."
}

Example Input (MCP Mode):

{
  "mode": "mcp"
}

Output Format (Normal Mode)

The actor returns a JSON object containing the execution details and the AI's result.

{
  "mode": "normal",
  "aiProvider": "openai",
  "pdfUrl": "https://example.com/document.pdf",
  "charactersExtracted": 12500,
  "aiResult": "The AI-generated summary and analysis appears here..."
}

In MCP mode, results are streamed directly to the connected AI agent via the WebSocket connection.

Project Structure

The main logic is in main.js. Key modules within /src include:
* orchestrator/: Manages the execution flow.
* connectors/: Contains adapters for each AI provider (OpenAI, Google, Anthropic).
* mcp/: Houses the WebSocket server and tool handlers for MCP mode.
* utils/: Provides utilities for PDF processing, AI calls, and file management.