AI Blog Agent

A local AI-powered research agent that searches the web, fetches real content, and generates grounded answers using a local Ollama model — exposed as an MCP (Model Context Protocol) tool for Claude Desktop.

---

🧠 How It Works

Query → Summarize → Generate Search Query → Tavily Search → Fetch Docs → Grounded Answer

Step	Method	Description
1	summarize()	Expands the query into context using local Ollama model
2	make_search_query()	Condenses query into a short search string (under 400 chars)
3	search_web()	Searches the web using Tavily API
4	fetch_docs()	Scrapes and cleans text from URLs (skips blocked domains)
5	uni_function()	Combines all steps and generates a final grounded answer

---

🚀 Features

• 🔍 Real-time web search via Tavily API
• 🧹 Automatic content cleaning (removes scripts, navbars, footers)
• 🚫 Blocked domain filtering (Medium, YouTube, Twitter, Reddit)
• 🤖 Local LLM inference via Ollama
• 🔌 MCP tool integration for Claude Desktop
• 🧪 Test mode for quick pipeline validation

---

📦 Requirements

• Python 3.10+
• Ollama running locally with gpt-oss:120b-cloud model
• Tavily API key — get one at app.tavily.com

---

🛠️ Installation

1. Clone the repo:

git clone https://github.com/BhavinXAgheda/AI_Blog_MCP_Agent.git
cd AI_Blog_MCP_Agent

2. Create and activate virtual environment:

python -m venv venv
source venv/bin/activate        # Mac/Linux
venv\Scripts\activate           # Windows

3. Install dependencies:

pip install fastmcp ollama tavily-python requests beautifulsoup4 python-dotenv

4. Create .env file:

cp .env.example .env

Then edit .env and add your Tavily API key:

TAVILY_API_KEY=your-tavily-api-key-here

---

▶️ Usage

Test the pipeline:

python test.py test

Start as MCP server:

python test.py

---

🔌 Claude Desktop Integration

Add this to your claude_desktop_config.json:

Mac: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "blog-agent": {
      "command": "/path/to/venv/bin/python",
      "args": ["/path/to/AI_Blog_MCP_Agent/test.py"]
    }
  }
}

Replace the paths with your actual venv and project paths, then restart Claude Desktop.

You can then ask Claude:

"Research the latest AI news in March 2026"

And it will call your local agent to search, fetch, and answer using live web data.

---

📁 Project Structure

AI_Blog_MCP_Agent/
├── test.py           # Main agent + MCP server
├── .env              # Your API keys (never committed)
├── .env.example      # Template for environment variables
├── .gitignore        # Ignores .env, venv, __pycache__
└── README.md         # This file

---

🔐 Environment Variables

Variable	Description
TAVILY_API_KEY	Your Tavily search API key

---

🚫 Blocked Domains

The following domains are skipped during doc fetching (paywalled or JS-heavy):

• medium.com
• youtube.com
• twitter.com
• reddit.com

You can extend the BLOCKED_DOMAINS list in test.py as needed.

---

🧪 Example Output

Query:        How do I handle file uploads in Next.js 14?
Search Query: Next.js 14 file upload handling
Summary:      The user is asking for a guide on implementing file upload...
URLs:         ['https://oneuptime.com/blog/...', 'https://dev.to/...']
Docs fetched: 2
Answer:       ## Handling File Uploads in Next.js 14 ...

---

📄 License

MIT License — feel free to use, modify, and distribute.

---

🙌 Built With

• FastMCP — MCP server framework
• Ollama — Local LLM inference
• Tavily — Web search API
• BeautifulSoup4 — HTML parsing