​

Custom Tools via MCP

​

What is MCP?

• Dynamic tool loading - Load tools from external servers
• Service integration - Connect to any HTTP service
• Custom functionality - Add specialized capabilities
• Multi-tenant support - User-isolated tool access

​

How MCP Tools Work

1. Request - Agent generates tool call request
2. MCP Routing - MCP adapter routes to appropriate MCP server
3. Execution - MCP server executes the tool
4. Response - Results returned through MCP adapter
5. Output - Agent uses tool results in response

​

MCP Tool Types

​

Built-in MCP Support

• Use published agents as tools
• Task delegation and composition
• Multi-agent workflows

• HTTP endpoint
• JSON request/response format
• Tool discovery endpoint

​

Using MCP Tools

​

Enabling MCP Tools

1. Go to Build page
2. Create or edit an agent
3. In the Tools section:
   • Enable mcp tool category
   • Agent can now access MCP tools

​

Configuration

• MCP servers registered in Xagent settings
• Tools automatically available to agents with MCP enabled
• User permissions control tool access

​

Agent Usage

• Discover available MCP tools
• Call tools with appropriate parameters
• Handle tool results
• Chain MCP tools with other tools

​

Building MCP Servers

​

MCP Server Requirements

GET /tools

{
  "tools": [
    {
      "name": "my_custom_tool",
      "description": "Description of what the tool does",
      "parameters": {
        "type": "object",
        "properties": {
          "param1": {
            "type": "string",
            "description": "Parameter description"
          }
        },
        "required": ["param1"]
      }
    }
  ]
}

POST /tools/{tool_name}

{
  "param1": "value1"
}

{
  "result": "Tool execution result",
  "status": "success"
}

​

Example MCP Server

from fastapi import FastAPI
from pydantic import BaseModel

app = FastAPI()

class ToolRequest(BaseModel):
    query: str

@app.get("/tools")
async def list_tools():
    return {
        "tools": [
            {
                "name": "company_database_search",
                "description": "Search internal company database",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "query": {
                            "type": "string",
                            "description": "Search query"
                        }
                    },
                    "required": ["query"]
                }
            }
        ]
    }

@app.post("/tools/company_database_search")
async def search_database(request: ToolRequest):
    # Implement your custom logic here
    result = query_company_database(request.query)
    return {
        "result": result,
        "status": "success"
    }

def query_company_database(query: str) -> str:
    # Your custom implementation
    return f"Database results for: {query}"

​

Deploying MCP Servers

1. Internal Server - Host within your network
   • Fastest performance
   • Full control
   • Direct database access
2. Cloud Service - Deploy to cloud (AWS, GCP, Azure)
   • Scalable
   • Accessible from anywhere
   • Managed infrastructure
3. Serverless - Use AWS Lambda, Google Cloud Functions
   • Pay-per-use
   • Auto-scaling
   • No server management

​

Security Considerations

​

Authentication

• API keys in request headers
• OAuth2 tokens
• Mutual TLS

POST /tools/my_tool
Authorization: Bearer YOUR_API_KEY

​

Authorization

• MCP tools respect user permissions
• Tools filtered by allowed_collections
• Admin-controlled tool availability

​

Best Practices

• Validate inputs - Sanitize all parameters
• Rate limiting - Prevent abuse
• Logging - Track tool usage
• Error handling - Return clear error messages
• Timeouts - Set appropriate timeouts

​

Troubleshooting

​

MCP Tools Not Available

• MCP server is running and accessible
• Server registered in Xagent settings
• Agent has mcp tool category enabled
• User has permission to access tools

​

Tool Calls Failing

• Tool endpoint is correct
• Parameters match expected schema
• Authentication is configured
• Network connectivity to MCP server

​

Slow Tool Execution

• Check MCP server performance
• Reduce response payload size
• Use caching for frequently accessed data
• Consider server location/latency

​

Advanced Features

​

Tool Composition

1. User request
2. Agent calls custom MCP tool (e.g., query database)
3. Uses results with built-in tools (e.g., Python executor)
4. Generates final response

​

Agent-as-Tool

• Create specialized agents
• Publish them
• Other agents can use them as tools
• Enables multi-agent workflows

​

Dynamic Tool Discovery

• Add/remove tools dynamically
• Update tool descriptions
• Change parameters
• Xagent discovers changes automatically

​

Resources

• MCP Specification - Official protocol documentation
• Built-in Tools - Reference for Xagent’s built-in tools
• Building Agents - Configure tools in your agents

​

Next Steps

• Tools Overview - Learn about tool categories
• Built-in Tools - Reference for all built-in tools
• Building Agents - Create tool-enabled agents