Universal MCP Server

Universal MCP Server

A model-agnostic Model Context Protocol (MCP) server implementation that works with any compatible AI model or client, not just Claude Desktop.

🎯 Project Goals

• Universal Compatibility: Works with any model that supports MCP (Claude, local models via Hugging Face, OpenAI, etc.)
• Simple Architecture: Clean, from-scratch implementation following official MCP specification
• Extensible Tools: Easy to add new tools and capabilities
• Learning-Focused: Well-documented code to understand MCP internals

📋 Project Scope

Phase 1: Core MCP Server

• JSON-RPC 2.0 over stdio communication
• Basic MCP protocol methods (initialize, tools/list, tools/call)
• File reading tool for specified directories
• Error handling and validation
• Configuration via command line/config file

Phase 2: Tool Expansion

• File writing capabilities
• Directory listing and navigation
• Text processing tools (search, replace, etc.)
• System information tools
• Custom tool plugin system

Phase 3: Multi-Model Client

• Generic MCP client library
• Hugging Face model integration
• OpenAI API integration
• Local model support (Ollama, etc.)
• Web interface for testing

🏗️ Architecture

┌─────────────────┐    JSON-RPC     ┌─────────────────┐
│   AI Model      │ ◄──────────────► │   MCP Server    │
│ (Any Provider)  │     (stdio)     │   (Python)      │
└─────────────────┘                 └─────────────────┘
                                            │
                                            ▼
                                    ┌─────────────────┐
                                    │     Tools       │
                                    │ • File Reader   │
                                    │ • File Writer   │
                                    │ • Directory Ops │
                                    └─────────────────┘

🚀 Quick Start

Running the MCP Server

# Install dependencies
pip install -r requirements.txt

# Run the server (communicates via stdio)
python mcp_server.py --allowed-paths ./data ./documents

# Test with a simple echo
echo '{"jsonrpc":"2.0","method":"tools/list","id":1}' | python mcp_server.py

Integrating with Models

Claude Desktop

{
  "mcpServers": {
    "file-tools": {
      "command": "python",
      "args": ["path/to/mcp_server.py", "--allowed-paths", "./data"]
    }
  }
}

Hugging Face Models

from mcp_client import MCPClient
from transformers import pipeline

# Initialize your model
model = pipeline("text-generation", model="microsoft/DialoGPT-medium")

# Connect to MCP server
mcp_client = MCPClient("python mcp_server.py")

# Use tools through the model
response = model("Can you read the file data/example.txt?")
tool_result = mcp_client.call_tool("read_file", {"path": "data/example.txt"})

🛠️ Available Tools

File Operations

• read_file: Read contents of a file within allowed paths
• list_directory: List files and folders in a directory
• file_info: Get file metadata (size, modified date, etc.)

Planned Tools

• write_file: Write content to files
• search_files: Search for text within files
• execute_command: Run system commands (with safety restrictions)

📁 Project Structure

universal-mcp-server/
├── mcp_server.py           # Main MCP server implementation
├── mcp_client.py           # Generic client for any model
├── tools/
│   ├── __init__.py
│   ├── file_tools.py       # File operation tools
│   └── system_tools.py     # System information tools
├── examples/
│   ├── huggingface_client.py
│   ├── openai_client.py
│   └── test_tools.py
├── config/
│   └── server_config.yaml
├── requirements.txt
└── README.md

🔧 Configuration

Server Configuration (config/server_config.yaml)

server:
  name: "Universal File Tools"
  version: "1.0.0"
  
security:
  allowed_paths:
    - "./data"
    - "./documents"
  max_file_size: "10MB"
  
tools:
  file_reader:
    enabled: true
  file_writer:
    enabled: false  # Disabled by default for security

Command Line Options

python mcp_server.py \
  --config config/server_config.yaml \
  --allowed-paths ./data ./docs \
  --max-file-size 5MB \
  --log-level INFO

🧪 Testing

Unit Tests

python -m pytest tests/

Manual Testing

# Test tool listing
echo '{"jsonrpc":"2.0","method":"tools/list","id":1}' | python mcp_server.py

# Test file reading
echo '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"read_file","arguments":{"path":"data/test.txt"}},"id":2}' | python mcp_server.py

Integration Tests

# Test with different models
python examples/test_huggingface.py
python examples/test_openai.py

🔐 Security Considerations

• Path Restrictions: Only allow file access within specified directories
• File Size Limits: Prevent reading of extremely large files
• Input Validation: Sanitize all tool parameters
• Command Execution: Disabled by default, whitelist approach when enabled

🤝 Contributing

1. Follow the official MCP specification
2. Add tests for new tools
3. Update documentation
4. Ensure compatibility across different model providers

📚 Resources

• Official MCP Specification
• MCP Python SDK
• JSON-RPC 2.0 Specification

🎓 Learning Outcomes

By building this project, you'll understand:

• How MCP protocol works under the hood
• JSON-RPC communication patterns
• Building model-agnostic AI tool interfaces
• Security considerations for AI tool access
• Integrating with various AI model providers

Next Steps: Start with mcp_server.py implementing basic file reading, then expand to multiple tools and model integrations!