MCP Hub

MCP Hub Project - Deep Research & Code Assistant

Overview

The MCP (Model Context Protocol) Hub is a sophisticated research and code assistant built using Gradio's MCP server functionality. This project demonstrates how to build a workflow of connected AI agents that work together to provide deep research capabilities and generate executable Python code.

The system orchestrates an 8-step deep research and code generation workflow:

1. Question Enhancement: Breaks down a user's original query into three distinct sub-questions
2. Web Search: Conducts web searches for each sub-question using Tavily API
3. LLM Summarization: Summarizes search results for each sub-question using Nebius LLMs
4. Citation Formatting: Extracts and formats citations from web search results
5. Result Combination: Merges all summaries into a comprehensive grounded context
6. Code Generation: Creates Python code based on the research findings using Qwen2.5-Coder-32B-Instruct-fast
7. Code Execution: Runs the generated code in a Modal sandbox environment
8. Final Summary: Provides a natural language summary of the entire process

Features

• MCP Server Implementation: Built on Gradio's MCP server capabilities for seamless agent communication
• Multi-Agent Architecture: Demonstrates how to build interconnected agent services
• Real-time Web Search: Integration with Tavily API for up-to-date information
• LLM Processing: Uses Nebius (OpenAI-compatible) models for text processing
• Structured Workflow: Showcases a sophisticated multi-step AI research process
• Citation Generation: Automatically formats APA-style citations from web sources
• Code Generation: Creates executable Python code based on research findings
• Code Execution: Runs generated code in a Modal sandbox environment
• Final Summary: Provides a natural language summary of the entire process

Prerequisites

• Python 3.12+
• API keys for:
  • Nebius API
  • Tavily API
• Modal account (for code execution in sandbox)

Installation

1. Clone this repository
2. Create a virtual environment (recommended)

3. Install dependencies:

4. Create a .env file with the following content:

Usage

Run the main application:

This will launch the Gradio interface at http://127.0.0.1:7860/

The MCP schema will be available at http://127.0.0.1:7860/gradio_api/mcp/schema

Available Agents

The project includes several agent services:

1. Question Enhancer: Splits a request into three sub-questions using Qwen3-4B-fast
2. Web Search Agent: Performs web searches via Tavily API (top-3 results)
3. LLM Processor: Processes text with Nebius LLMs (Meta-Llama-3.1-8B-Instruct) for summarization, reasoning, or keyword extraction
4. Citation Formatter: Extracts URLs and formats them as APA-style citations
5. Code Generator: Creates Python code snippets based on research context using Qwen2.5-Coder-32B-Instruct-fast
6. Code Runner: Executes Python code in a Modal sandbox environment
7. Orchestrator: Coordinates all agents in a cohesive workflow

Tutorial Scripts

The tutorial_scripts/ directory contains example Gradio applications and code samples for learning:

• simple_app.py: A basic Gradio interface
• letter_count.py: A simple letter counting example
• predict_letter_count.py: Example of letter counting prediction
• modal_inference.py: Demonstrates using Modal for code execution
• nebius_inference.py: Shows how to use Nebius API for inference
• nebius_tool_calling.py: Example of tool calling with Nebius models
• Gradio Cheat Sheet.md: Quick reference for Gradio features and usage

MCP Implementation Details

This project demonstrates how to:

• Create MCP-compatible function definitions with proper typing and docstrings
• Launch a Gradio app as an MCP server (mcp_server=True)
• Structure a multi-agent workflow
• Pass data between agents in a structured format
• Execute code safely in a sandbox environment

Example Workflow

1. A user submits a high-level request like "Write Python code to analyze sentiment from Twitter data"
2. The system breaks this into three sub-questions (e.g., about Twitter APIs, sentiment analysis techniques, and Python libraries)
3. For each sub-question, it:
   • Performs a web search using Tavily
   • Summarizes the search results
   • Extracts citations from URLs
4. The sub-summaries are combined into a comprehensive grounded context
5. Based on this context, Python code is generated
6. The code is executed in a Modal sandbox
7. The user receives the final summary, generated code, execution output, and citations

License

[Your license information here]

Contributing

[Your contribution guidelines here]

Acknowledgments

• Gradio for providing the MCP server functionality
• Nebius for LLM capabilities
• Tavily for web search capabilities