MCP Server with Dynamic Per-User Tool Generation

A Model Context Protocol (MCP) server implementation using FastMCP that dynamically generates tools based on user permissions from an external API.

Features

• Dynamic Tool Generation: Tools are generated at runtime based on user-specific schemas from an external API
• Per-User Permissions: Each authenticated user receives tools tailored to their permissions
• Intelligent Caching: Both API schemas and generated tools are cached with TTL for performance
• Graceful Degradation: Falls back to static tools if API is unavailable
• Comprehensive Logging: Detailed logging at INFO, DEBUG, WARNING, and ERROR levels
• Thread-Safe: Concurrent requests from multiple users are handled safely

Architecture

Components

1. mcp_server.py - Main server entry point

   • Initializes FastMCP server
   • Registers static tools (add, echo, multiply)
   • Sets up dynamic tool middleware

2. api_client.py - External API integration

   • FormSchemaClient: Fetches and parses schemas from external API
   • FormSchemaCache: Caches API responses with TTL
   • Converts API field definitions to JSON Schema format

3. dynamic_tool_manager.py - Tool lifecycle management

   • DynamicToolManager: Manages per-user tool storage
   • Thread-safe caching with TTL
   • Cache statistics and invalidation

4. tool_function_factory.py - Dynamic function generation

   • Creates typed async functions from JSON schemas
   • Generates proper inspect.Signature for FastMCP validation
   • Maps JSON Schema types to Python types

5. tool_execution_handler.py - Tool execution backend

   • ToolExecutionRouter: Routes tool calls to appropriate handlers
   • Specialized handlers for different tool types
   • Structured error responses

6. dynamic_tool_middleware.py - FastMCP middleware

   • Intercepts list/tools requests
   • Extracts authentication tokens
   • Generates and caches user-specific tools
   • Returns combined static + dynamic tools

Installation

1. Install dependencies:

   pip install -r requirements.txt
   

2. Configure the API URL: Edit mcp_server.py and set the FORM_SCHEMA_API_URL to your external API endpoint:

   FORM_SCHEMA_API_URL = "http://172.16.11.131/api/module/request/form"
   

Usage

Starting the Server

python mcp_server.py

The server will start on http://127.0.0.1:9092/mcp

Authentication

The server expects authentication via HTTP Authorization header:

Authorization: Bearer <your-token>

The token is used to:

1. Fetch user-specific schema from the external API
2. Generate tools with fields the user has permission to access
3. Cache tools for subsequent requests

Available Tools

Static Tools (Always Available)

• add(a: int, b: int) -> int - Adds two numbers
• echo(message: str) -> str - Echoes a message
• multiply(a: int, b: int) -> int - Multiplies two numbers

Dynamic Tools (User-Specific)

• create_request(...) - Creates a request with fields based on user permissions
  • Parameters are dynamically generated from the external API schema
  • Each user sees different parameters based on their permissions
  • Required fields, enums, and types are enforced

How It Works

Request Flow

1. Client sends list/tools request with Authorization: Bearer <token>

2. Middleware intercepts the request:

   • Extracts auth token from header
   • Checks if tools are cached for this user
   • If cached: Returns cached tools
   • If not cached: Proceeds to generation

3. Tool Generation:

   • Fetches schema from external API with user's token
   • Parses API response into JSON Schema
   • Creates typed async function with proper signature
   • Converts function to FastMCP Tool object
   • Caches the tool for future requests

4. Response:

   • Returns list of static tools + dynamic tools
   • Client sees tools specific to their permissions

Tool Execution Flow

1. Client calls create_request tool with arguments

2. FastMCP validates arguments against the generated function signature

3. Tool execution handler:

   • Receives validated arguments
   • Processes the request
   • Returns structured response with ID, timestamp, status

Configuration

Cache TTL

Both schema cache and tool cache use 5-minute TTL by default. To change:

# In mcp_server.py
schema_client = FormSchemaClient(
    api_url=FORM_SCHEMA_API_URL,
    cache_ttl=300,  # Change this value (seconds)
    verbose=True
)
tool_manager = DynamicToolManager(cache_ttl_seconds=300)  # Change this value

Logging Level

To change logging verbosity:

# In mcp_server.py
logging.basicConfig(
    level=logging.DEBUG,  # Change to DEBUG, INFO, WARNING, or ERROR
    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s'
)

Testing

See TEST_SCENARIOS.md for comprehensive test scenarios including:

• User-specific tool generation
• Caching behavior
• Tool execution
• Error handling
• Validation enforcement

API Schema Format

The external API should return a response with this structure:

{
  "fieldList": [
    {
      "name": "Subject",
      "paramName": "subject",
      "type": "TextFieldRest",
      "required": true,
      "description": "Request subject",
      "groupIds": [1, 2],
      "hidden": false,
      "removed": false,
      "inActive": false
    }
  ]
}

Supported field types:

• TextFieldRest → string
• NumberFieldRest → number
• DropDownFieldRest → string with enum
• MultiSelectDropDownFieldRest → array of strings
• And more (see api_client.py for full mapping)

Security Considerations

• Authentication tokens are truncated in logs and responses
• Tokens are validated on each request
• No token = static tools only (graceful degradation)
• FastMCP validates all tool arguments before execution

Troubleshooting

Server won't start

• Check that all dependencies are installed: pip install -r requirements.txt
• Verify Python version is 3.12+
• Check logs for import errors

Tools not appearing

• Verify Authorization header is present and correctly formatted
• Check server logs for API fetch errors
• Ensure external API is accessible
• Verify user has permissions in the external system

Cache not updating

• Default TTL is 5 minutes
• Manually clear cache: Call tool_manager.clear_all_tools() or restart server
• Check logs for cache hit/miss messages

License

[Your License Here]

Contributing

[Your Contributing Guidelines Here]