Create New MCP Server

You are helping the user create a new MCP (Model Context Protocol) server from scratch.

Philosophy

"This is not a code walkthrough. The true value lies in the problem you're solving, how you think about solving it, and empathizing with the client (the agent consuming your server)."

Workflow

Phase 1: Understand the Problem

Ask the user:

•What problem are you solving? What will the agent help users accomplish?
•Who is the end user? (e.g., financial advisors, developers, healthcare workers)
•What data/systems will the agent access? (APIs, databases, files)

Phase 2: Choose Transport

Ask the user which transport they prefer:

Transport	Best For	Description
Streamable HTTP (Recommended)	Web deployment, ChatGPT Apps, remote access	Server runs as HTTP service, clients connect via URL
stdio	Local development, Claude Desktop	Server communicates via stdin/stdout

Streamable HTTP advantages:

•Easier testing with MCP Inspector (connect via URL)
•Ready for cloud deployment (Cloud Run, Render, Fly.io)
•Supports multiple concurrent clients
•Required for ChatGPT Apps integration

stdio advantages:

•Simpler setup for local-only use
•Direct integration with Claude Desktop
•No port management needed

Phase 3: Choose Architecture Pattern

Present these options based on their needs:

Pattern	Best For	Complexity
Basic	Learning, simple tools, quick prototypes	Low
Tool-Heavy	Data analysis, many tools (like financial research)	Medium
Widget-Enabled	Rich UI output, ChatGPT Apps, visualizations	Medium-High
Database-Backed	User auth, state persistence, multi-tenant	High

Phase 4: Identify True Tasks

Guide the user to think about user intents, NOT API endpoints:

code

User Intent → Outcome → Tool Design
"What's AAPL's price?" → Complete stock summary → get_stock_summary(ticker)
"Compare these stocks" → Side-by-side analysis → compare_stocks(tickers[])

Anti-pattern to avoid:

•Creating get_task_ids() + get_task_details() when you could have search_tasks(include_details=true)

Phase 5: Define Initial Tools

For each tool, capture:

•Name (kebab-case): get-stock-summary, search-tasks
•Description: When should the agent use this? (not just what it does)
•Inputs: What parameters? Which are required?
•Output: What does a complete result look like?

Phase 6: Generate Server

Based on the chosen pattern and transport, use the mcp-builder:mcp-architect agent to generate:

For Python (FastMCP) with Streamable HTTP:

code

{server-name}/
├── server.py              # FastMCP server with HTTP transport
├── requirements.txt       # Python dependencies
├── Dockerfile            # Container config
├── .env.example          # Environment template
├── setup.sh              # One-time setup script
├── START.sh              # Multi-mode startup script
└── src/                  # Business logic (if tool-heavy)
    └── {domain}.py       # Domain-specific modules

For TypeScript (MCP SDK):

code

{server-name}/
├── server/
│   └── index.ts          # MCP server
├── package.json
├── tsconfig.json
├── Dockerfile
├── setup.sh
├── START.sh
└── .env.example

Phase 7: Startup Scripts

Generate START.sh with multiple modes:

bash

#!/bin/bash
# START.sh - Multi-mode MCP server launcher

case "${1:-}" in
  --dev)
    # Development mode with auto-reload
    echo "Starting in dev mode..."
    python server.py
    ;;
  --inspector)
    # Start server + open MCP Inspector
    echo "Starting server and MCP Inspector..."
    python server.py &
    SERVER_PID=$!
    sleep 2
    npx @modelcontextprotocol/inspector
    kill $SERVER_PID
    ;;
  --test)
    # Run tests
    pytest tests/
    ;;
  *)
    # Production mode
    python server.py
    ;;
esac

Phase 8: Verify Setup

After generation:

•Run setup: ./setup.sh
•Start in dev mode: ./START.sh --dev
•
Test with Inspector: ./START.sh --inspector
- •Or manually: npx @modelcontextprotocol/inspector
- •Select "Streamable HTTP" transport
- •Enter URL: http://localhost:8000/mcp

Server Naming

Use kebab-case: stock-analysis-server, patient-records-server, task-manager-server

Language Selection

Python (Recommended for most cases):

•FastMCP makes tool creation simple
•Great for data processing, API integrations
•Most MCP examples are in Python

TypeScript (For widget-enabled servers):

•Required for ChatGPT Apps with inline widgets
•Better if you're building React UI components

State Management

Create .mcp-builder/state.json to track progress:

json

{
  "serverName": "stock-analysis-server",
  "pattern": "tool-heavy",
  "language": "python",
  "tools": [
    { "name": "get-stock-summary", "status": "implemented" }
  ],
  "resources": [],
  "createdAt": "2025-01-15T00:00:00Z"
}

Next Steps

After scaffolding, suggest:

•/mcp-builder:add-tool - Add more tools
•/mcp-builder:validate - Check best practices
•/mcp-builder:test - Test with MCP Inspector