Unified Research Assistant

A comprehensive research tool that integrates multiple MCP servers with Anthropic's Claude for intelligent information gathering and organization.

Features

• Pure MCP Integration: Connects to Brave, Puppeteer, and Notion MCP servers with full AI-driven tool selection
• Anthropic Tool Calling: Uses Claude to intelligently call Brave search and Puppeteer scraping tools
• Content Scraping: Extracts detailed information from web pages using intelligent tool selection
• Vector Storage: Stores research data in a Chroma vector database with semantic search capabilities
• Summarization: Generates comprehensive summaries of research findings
• Notion Integration: Organizes and saves research results in Notion

Prerequisites

• Python 3.10+
• Node.js 16+ (for JavaScript-based MCP servers)
• Anthropic API key (Claude 3 Sonnet model access required)
• Brave Search API key
• Notion API token and workspace access

Detailed Setup

1. Environment Setup

1. Clone this repository
2. Create and activate a virtual environment:

   python -m venv .venv
   source .venv/bin/activate  # On Windows: .venv\Scripts\activate

3. Install dependencies:

   pip install -r requirements.txt

2. API Keys Configuration

Create a .env file in the project root with the following variables:

# Required API Keys
ANTHROPIC_API_KEY=your_anthropic_api_key_here
BRAVE_API_KEY=your_brave_api_key_here
NOTION_TOKEN=your_notion_token_here

# Optional Settings
NOTION_PAGE_ID=your_notion_page_id_here
CHROMA_PERSIST_DIRECTORY=research_db

• ANTHROPIC_API_KEY: Get from Anthropic Console
• BRAVE_API_KEY: Get from Brave Search API
• NOTION_TOKEN: Get from Notion Integrations
• NOTION_PAGE_ID (optional): ID of the Notion page where research will be saved

3. ChromaDB Setup

The application uses ChromaDB for vector storage with semantic search capabilities. The system uses:

• PersistentClient: For reliable on-disk storage
• SentenceTransformer Embeddings: Using the "all-MiniLM-L6-v2" model for high-quality embeddings
• Automatic Data Persistence: Changes are automatically saved to disk

By default, data is stored in a local directory named research_db. You can modify this by setting the CHROMA_PERSIST_DIRECTORY environment variable.

# Optional: Install additional ChromaDB dependencies for better performance
pip install chromadb[all]

First-time setup may take longer as the sentence transformer model is downloaded.

4. MCP Server Setup

You'll need to set up the following MCP servers:

1. Brave Search MCP Server:

   npx -y @modelcontextprotocol/server-brave-search

   This will initialize the server and provide a path to the server script.

2. Puppeteer MCP Server:

   npx -y @modelcontextprotocol/server-puppeteer

3. Notion MCP Server:

   npx -y @modelcontextprotocol/server-notion

Make note of the paths to each server script, as you'll need them when running the application.

Usage

Run the application with:

python main.py "your research query" \
    --brave-server path/to/brave/server.js \
    --puppeteer-server path/to/puppeteer/server.js \
    --notion-server path/to/notion/server.js

Example:

python main.py "The impact of quantum computing on cybersecurity" \
    --brave-server ./node_modules/.bin/mcp-server-brave-search \
    --puppeteer-server ./node_modules/.bin/mcp-server-puppeteer \
    --notion-server ./node_modules/.bin/mcp-server-notion

How It Works

1. MCP Server Connection:

   • Connects to each MCP server
   • Lists available tools and their parameters

2. Anthropic Tool Calling for Search:

   • Converts MCP tool schemas to Anthropic tool format
   • Uses Claude to intelligently select and call Brave search tools
   • No fallbacks - Claude is fully responsible for tool selection

3. Anthropic Tool Calling for Scraping:

   • Analyzes each URL to determine the best scraping approach
   • Uses Claude to select appropriate Puppeteer tools based on the content
   • Strictly follows MCP philosophy - only scrapes URLs where Claude selects a tool

4. Content Processing:

   • Stores all data in the ChromaDB vector database with semantic embeddings
   • Generates a comprehensive summary using Claude

5. Result Organization:

   • Saves organized research results to Notion
   • Maintains a persistent local database of all research data for future queries

MCP Implementation Details

The application follows a pure Model Context Protocol (MCP) specification:

1. Server Discovery: Connects to MCP servers and discovers available tools
2. Tool Schema Conversion: Converts MCP tool schemas to Anthropic tool format
3. AI-Driven Tool Selection: Claude has full responsibility for selecting appropriate tools
4. Tool Execution: Executes MCP tool calls based on Claude's recommendations without fallbacks

Advanced Configuration

ChromaDB Configuration

The application uses ChromaDB's PersistentClient with the sentence-transformers embedding model. You can customize these settings:

• Storage Path: Set CHROMA_PERSIST_DIRECTORY in your .env file
• Embedding Model: The code uses "all-MiniLM-L6-v2" by default, but you can modify it for different models
• Collection Configuration: Metadata and settings can be adjusted in the code

For production use, consider setting up a dedicated directory on persistent storage.

Notion Integration

For better Notion integration:

1. Create a dedicated Notion integration at Notion Integrations
2. Share a specific page or database with your integration
3. Copy the page ID and set it as NOTION_PAGE_ID in your .env file

Troubleshooting

Connection Issues

• Problem: Unable to connect to MCP servers
• Solution: Ensure servers are running and paths are correct. Try running the server commands in a separate terminal to verify they work.

API Key Errors

• Problem: "Missing required environment variables" error
• Solution: Check that all required API keys are set in your .env file with the correct variable names.

Tool Calling Failures

• Problem: Claude is not selecting tools
• Solution: Ensure you're using Anthropic API version 0.18.0+ and Claude 3 Sonnet or higher model.

Missing Data

• Problem: URLs aren't being scraped
• Solution: Check the logs to see if Claude is selecting tools. If not, try updating the prompts in the code.

ChromaDB Errors

• Problem: "Failed to initialize ChromaDB" error
• Solution: Ensure the directory specified in CHROMA_PERSIST_DIRECTORY is writable.
• Problem: Slow first-time startup
• Solution: This is normal as the sentence transformer model downloads. Subsequent runs will be faster.