Inori MCP

🚀 A Very Simple MCP Server

This project implements a simplest MCP server using Go and Python respectively, and it can be invoked by an MCP client. This project doesn't rely on any external APIs, allowing you to easily understand a simple MCP workflow.

🚀 Quick Start

Scenario Assumption

There are 4 shelves of books numbered A, B, C, and D. I need a large model to help me find the books corresponding to the given numbers.

Implementation Files

• bookSearch.go: Go implementation of the MCP tool.
• MCPServer_book_search.py: Python implementation of the MCP tool.
• README.md: Explanation of the Python workflow.

Writing Method

Specify MCP Input Parameters

type BookShelfParams struct {
    BookShelfId string `json:"book_shelf_id" jsonschema:"required,description=The ID of the book shelf"`
}

Core Logic

func bookSearchHandler(ctx context.Context, args BookShelfParams) (*mcp.CallToolResult, error) {
    // Logic implementation, can be replaced with any external API call.
    return mcp.NewToolResultText("Return result:xxxx"), nil
}

Define Tool Name and Description

var BookSearchTool = InoriMCP.MustTool(
    "book-search",
    "Book search tool",
    bookSearchHandler,
)

Entry Function

func AddBookSearchTools(mcp *server.MCPServer) {
    BookSearchTool.Register(mcp)
}

Add this tool in main.go:

tools.AddBookSearchTools(s)

Build command:

go build -o inori-mcp.exe ./cmd/inori-mcp

Start the MCP server:

mcp.exe

The startup methods of the MCP clients in Go and Python are different.

Debugging

Install the Node environment and use the inspector tool. First, connect to the MCP server, then input A, B, C, or D to check if it returns the book information in the bookshelf.

npx -y @modelcontextprotocol/inspector .\inori-mcp.exe

Start the Client to Interact with the MCP

Modify the startup script of the connect_to_server function on line 51 in MCPClient.py:

server_params = StdioServerParameters(
    command=<mcp.exe> path,
    args=[],
    env=None
)

You need to create a .env file from .env.sample and fill in the base URL of the OpenAI-compatible large model. It is recommended to use the QWQ - 32B model (DeepSeek series is not recommended).

The model needs to trigger ToolCall and summarize the results, and finally return the book information in the bookshelf.