MCP Client Service (Gateway)

Central gateway that discovers, registers, and routes tool calls to individual MCP tool servers.

• Tech: Python, FastAPI
• Port: 8000
• Auth: x-jeen-mcp-service-secret header

Endpoints

Method	Path	Auth	Description
GET	/mcp/v1/tools/list-tools	Secret header	List all available tools across all registered servers
POST	/mcp/v1/tools/call-tool	Secret header	Execute a tool by name with arguments

GET /mcp/v1/tools/list-tools

Returns an aggregated list of all tools from all registered MCP servers. Each tool includes its name, description, and input schema (JSON Schema).

POST /mcp/v1/tools/call-tool

Routes a tool call to the correct MCP server based on the tool name.

Request:

{
  "name": "search_documents",
  "arguments": {
    "queries": ["kubernetes pod scheduling"],
    "similarityTopK": 10
  }
}

Response: Tool-specific result.

How It Works

1. Startup: The ToolRegistry connects to every configured MCP server and fetches their tool lists
2. Registration: Each tool name is mapped to its source server in an in-memory registry
3. List tools: Returns all tools from the registry
4. Call tool: Looks up the tool name in the registry, finds the source server, and forwards the call

Configuration

Development (config.json)

[
    {
        "name": "demo-server",
        "url": "http://127.0.0.1:3001/mcp"
    },
    {
        "name": "mcp-rag-tool",
        "url": "http://127.0.0.1:3002/mcp"
    }
]

A JSON array where each entry has a name (server identifier) and url (MCP endpoint).

Production

Server configurations are fetched from llm-core via HTTP (GET /services-config/mcp-servers), not from the local config file. The LLM_CORE_URL environment variable must be set.

Transport

Uses the MCP Streamable HTTP transport to communicate with tool servers. The Python mcp SDK v1.8.1 handles connection management.