quiquietus/MCP-Server-Puch-AI

Web Analyzer MCP Server

Overview

The Web Analyzer MCP Server is a Model Context Protocol (MCP) server that processes web pages and answers user queries about them.
It is designed to:

• Fetch web pages from a given URL.
• Clean and convert the content into Markdown.
• Split the text into manageable chunks.
• Use a multi-agent pipeline with Google Gemini to:
  1. Triage Agent – Identify relevant chunks for the query.
  2. Summarization Agent – Extract key points from relevant chunks.
  3. Synthesizer Agent – Combine summaries into a final answer.

The server supports Bearer token authentication and integrates seamlessly with tools like Puch.

Features

• Multi-Agent Pipeline
  • Triage → Summarization → Final Synthesis.
• Gemini AI Integration
  • Uses Google’s gemini-2.0-flash-lite model (configurable).
• Markdown Conversion
  • Strips scripts, styles, headers, footers, and sidebars.

Requirements

Install dependencies from requirements.txt:

pip install -r requirements.txt

Python Version

Python 3.11+ is recommended.

Environment Variables

Create a .env file with:

AUTH_TOKEN=YOUR_AUTH_TOKEN
MY_NUMBER=YOUR_PHONE_NUMBER
GEMINI_API_KEY=YOUR_GEMINI_API_KEY
GEMINI_MODEL=gemini-2.0-flash-lite

Running the Server

Start the MCP server:

python demoServer.py

It will run on:

http://0.0.0.0:8086

Tools Provided

1. validate

Validates the server identity for Puch AI.

Response:
Returns the MY_NUMBER set in .env. (The MY_NUMBER entered is in the format {COUNTRY_CODE}{PHONE_NUMBER}. For example: If your phone number is +91-1234567890 then it will return 911234567890)

2. web_analyzer

Analyzes a webpage and answers questions about it.

Expected Input:

{
  "user_query": "Summarize https://en.wikipedia.org/wiki/Apple"
}

Process:

1. Extracts URL and question from input.
2. Fetches the page HTML.
3. Cleans and converts it to Markdown.
4. Runs the analysis pipeline.
5. Returns a final answer.

Notes

• The BATCH_SIZE and DELAY_BETWEEN_BATCHES variables in web_analyzer control request batching to Gemini to prevent rate-limit errors.
• Adjust GEMINI_MODEL in .env to use a different Gemini model.
• This implementation is stateless, i.e., each request is processed independently.