mcp-server

Simple MCP Demo

This demo shows how you can use Genkit's MCP package to expose an MCP Server that can be used by external consumers like Claude Desktop, VS Code with MCP, and other useful tools and clients.

Files

• server.go - MCP server with 2 tools: text_encode and hash_generate
• run-server.sh - Simple wrapper script for easy client setup
• client.go - Client that uses tools with Gemini AI

Two Ways to Run

Option A: Expose an MCP Server that works with Desktop Clients

Step 1: Test the server works

./run-server.sh

You should see: Starting server with tools: [text_encode hash_generate]

Step 2: Configure client.

Point your client to run that bash command to start the server. Make sure you restart the client to ensure the server is properly started.

Step 4: Test your assistant's connection to the MCP server with these quries:

• "Encode 'Hello World' as base64"
• "Generate an MD5 hash of 'password123'"
• "Decode this base64: SGVsbG8gV29ybGQ="

Option B: Genkit MCP Client spawns the MCP Server

First, change client() to main() in go/samples/mcp-server/client.go. Then follow the steps below:

export GOOGLE_AI_API_KEY=your_key
go run client.go

The GenkitMCPClient instance automatically spawns the server and makes tools available to use with client.GetActiveTools(). You can use these tools with any Generate method or ExecutablePrompt.

Alternatively, you can use the MCPManager for managing multiple servers:

manager, _ := mcp.NewMCPManager(mcp.MCPManagerOptions{
    Name: "my-app",
    MCPServers: []mcp.MCPServerConfig{
        {
            Name: "textUtils",
            Config: mcp.MCPClientOptions{
                Stdio: &mcp.StdioConfig{
                    Command: "go",
                    Args:    []string{"run", "server.go"},
                },
            },
        },
        {
            Name: "otherServer",
            Config: mcp.MCPClientOptions{
                URL: "http://localhost:8080/mcp",
            },
        },
    },
})

This example demonstrates the following workflow:

1. Client connects to server via MCP protocol
2. AI gets prompt: "Fetch content from URL and summarize it"
3. AI automatically calls fetch_url then summarize
4. Returns the summary

This option is useful when you have an genkit agent that needs access to utilities developed in the same project.