shantanujumde/mcp-client-server

MCP Server & Consumer

This workspace contains a minimal Model Context Protocol (MCP) server (mcp-server) and a companion consumer (mcp-consumer). The server exposes a simple get_length tool over Server-Sent Events (SSE). The consumer connects to the server, enumerates available tools, and calls the length tool through OpenAI's gpt-4o model using the ai SDK.

Project Layout

• mcp-server – Express + MCP server that registers the get_length tool and serves an SSE endpoint on port 4001.
• mcp-consumer – Node.js script that connects to the MCP server, lists the tools, and requests the length of "Hello World" via OpenAI.

Prerequisites

• Node.js 22 or later (tested with Node 22).
• npm 10 or later (bundled with Node 22).
• An OpenAI API key with access to the gpt-4o model.

Setup

1. Install dependencies for each package:

   cd mcp-server
   npm install
   
   cd ../mcp-consumer
   npm install
   

2. Create a .env file in mcp-consumer/ with the required environment variables (see below).

Environment Variables

Create mcp-consumer/.env containing the following key:

Optional overrides:

• MCP_SERVER_URL – Set to a custom SSE endpoint (defaults to http://localhost:4001/sse). If you use this variable, update index.ts accordingly.
• PORT – Change the HTTP port used by mcp-server (defaults to 4001). Update the consumer to match if you change it.

Running the Server

From mcp-server/:

• npm run dev – Start the MCP server with nodemon (auto-restarts on file changes).
• npm run build – Compile TypeScript to dist/.
• npm start – Run the built server from dist/index.js.

The server logs the listening URL and exposes:

• SSE endpoint: http://localhost:4001/sse
• Message relay endpoint: http://localhost:4001/messages

Running the Consumer

From mcp-consumer/:

• npm run dev – Start the consumer with ts-node and nodemon.
• npm run build – Compile TypeScript to dist/.
• npm start – Run the compiled consumer.

When the consumer runs it:

1. Connects to the SSE endpoint.
2. Lists available MCP tools in the console.
3. Uses gpt-4o to call the get_length tool and prints the tool output.
4. Closes the MCP client gracefully.

Development Notes

• TypeScript configs in each package emit CommonJS output under dist/.
• The server currently registers a single tool. Add additional tools via server.tool(...) in mcp-server/index.ts.
• The consumer invokes generateText with toolChoice: "required", forcing the model to use MCP tools for each request.

Troubleshooting

• Connection errors – Ensure the server is running on http://localhost:4001 before starting the consumer.
• 401/403 responses – Verify OPENAI_API_KEY is present and valid.
• Model mismatch – Adjust the model name in mcp-consumer/index.ts if your key does not support gpt-4o.