bun-mcp-server

A lightweight HTTP-based Model Context Protocol (MCP) server built with Bun.

Motivation

This project explores the MCP lifecycle through a clean, modular implementation built with Bun. By separating HTTP routing from JSON-RPC 2.0 handling, the architecture aims to remain transport-agnostic, providing a clear reference for how the protocol functions under the hood.

[!Note] This is an educational, work-in-progress implementation and does not cover the full MCP specification.

本プロジェクトは、Bun を使用したクリーンかつモジュール化された実装を通じて、MCP（Model Context Protocol）のライフサイクルを探求するものです。HTTPルーティングとJSON-RPC 2.0の処理を分離することでトランスポート層に依存しない設計を実現し、プロトコルの内部動作を理解するためのリファレンスとして機能します。

[!Note] 本プロジェクトは学習・研究用であり、MCP仕様のすべてを実装しているわけではありません。

You can read the full MCP specification here.

Features

• HTTP Transport — Serves MCP over a standard HTTP POST endpoint (/mcp), compatible with any MCP client that supports the HTTP transport.
• JSON-RPC 2.0 — Fully compliant request/response framing with proper error codes (-32700, -32600, -32601, -32602, -32603).
• Tool Registry — Pluggable tool system for easy registration and execution of tools.

Project Structure

bun-mcp-server/
├── src/
│   ├── index.js          # Entry point — starts the Bun HTTP server on port 3000
│   ├── lib/
│   │   └── utils.js      # Shared utilities (timestamps, CORS headers, logging, IP parsing)
│   ├── mcp/
│   │   ├── index.js      # Core MCP request handler (JSON-RPC dispatch)
│   │   └── tools.js      # Tool definitions and registration
│   └── routes/
│       └── index.js      # HTTP route definitions (/mcp, /.well-known/mcp, catch-all)
├── package.json
├── jsconfig.json
└── BRIEF.md              # Project brief / requirements

Prerequisites

• Bun v1.3+ installed

Getting Started

Install dependencies

bun install

Start the server

bun run dev

The server starts on http://localhost:3000 with file-watching enabled for live reload during development.

API Endpoints

POST /mcp

The main MCP endpoint. Accepts JSON-RPC 2.0 requests and dispatches them to the appropriate handler.

Supported methods:

GET /.well-known/mcp

This endpoint serves as the MCP Server Card. It provides essential metadata about the server—such as its name, description, supported transport types, and the base URL for the MCP endpoint—allowing clients to identify and connect to the service automatically.

While this mechanism is not currently part of the official Model Context Protocol (MCP) specification, it addresses a critical gap: enabling HTTP-based MCP servers to be discoverable via standard web patterns and simplifying the process for users to add them to their environments.

The return value will be mapped to the client's MCP server config file:

{
  "servers": {
    "weather-server": {
      "type": "http",
      "url": "http://localhost/mcp"
    }
  }
}

* /*

Catch-all route that returns a 404 Not Found JSON response for any unmatched path.

Usage Examples

Initialize the server

curl -X POST http://localhost:3000/mcp \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "initialize",
    "params": {
      "protocolVersion": "2024-11-05",
      "capabilities": {},
      "clientInfo": { "name": "test-client", "version": "1.0.0" }
    }
  }'

List available tools

curl -X POST http://localhost:3000/mcp \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 2,
    "method": "tools/list"
  }'

Call a tool

curl -X POST http://localhost:3000/mcp \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 3,
    "method": "tools/call",
    "params": {
      "name": "get_weather",
      "arguments": { "city": "Tokyo" }
    }
  }'

Discover the server

curl http://localhost:3000/.well-known/mcp

Live Testing

You can verify that the server is functioning correctly by connecting it to an MCP client, such as VS Code or AntiGravity.

VSCode

Follow these steps to configure and connect:

1. Create or open .vscode/mcp.json in your project directory.
2. Add the following entry:

{
  "servers": {
    "weather-server": {
      "url": "http://localhost:3000/mcp",
      "headers": {
        "Content-Type": "application/json"
      }
    }
  }
}

1. In the VS Code Chat panel, click the gear icon to open the Agent Customizations dialog.
2. Select MCP Servers from the sidebar. You should see weather-server listed under Workspace.
3. Right-click weather-server and select Start Server from the context menu to initialize the connection.

AntiGravity

Follow these steps to configure and connect:

1. Open Configuration: Navigate to Additional Options (...) and select MCP Servers.
2. Manage Servers: Click Manage MCP Servers, then select View raw config to open the mcp_config.json file.
3. Add Server: Add the following configuration:

{
  "mcpServers": {
    "weather-server": {
      "serverUrl": "http://localhost:3000/mcp",
      "headers": {
        "Content-Type": "application/json"
      }
    }
  }
}

1. Connect: Save the file, return to the Manage MCP Servers menu, and click Refresh.

[!Note] Ensure your MCP server is already running locally before connecting to either one. Upon success, you will see connection activity in your server’s console logs.

Once connected, confirm that tool discovery and execution are working by sending a prompt through your IDE's chat interface:

"Let's test the MCP server. Get the weather for Tokyo today."

If successful, the chat should trigger the tool and return the relevant data from your server.

Available Tools

The get_weather tool returns simulated data (random temperature, "Cloudy" condition) — it is included as a demo tool for testing the MCP lifecycle.

Adding New Tools

Register new tools in src/mcp/tools.js using the tool registry:

toolRegistry.register(
  "my_tool",
  {
    name: "my_tool",
    title: "My Tool",
    description: "Does something useful.",
    parameters: {
      type: "object",
      properties: {
        input: { type: "string", description: "Some input value" }
      },
      required: ["input"]
    }
  },
  async ({ input }) => {
    // Your tool logic here
    return { result: `Processed: ${input}` }
  }
)

The tool is automatically available via tools/list and executable via tools/call — no additional wiring needed.

Error Handling

The server returns standard JSON-RPC 2.0 error responses:

License

MIT © supershaneski