MCP Deployment and Testing for OpenAI Apps SDK

Overview

Deploying and testing MCP servers requires proper HTTPS setup, local development tools, and ChatGPT connector configuration. This skill covers the complete deployment workflow from local testing to production.

Local Development Setup

Running the Server

Python:

# Install dependencies
pip install mcp

# Run server
python server.py
# Server running on http://localhost:8000/mcp

TypeScript:

# Install dependencies
npm install @modelcontextprotocol/sdk express

# Run server
npx tsx server.ts
# Server running on http://localhost:8000/mcp

Exposing with ngrok

ChatGPT requires HTTPS. Use ngrok for local development:

# Install ngrok
brew install ngrok  # macOS
# or download from ngrok.com

# Expose local server
ngrok http 8000

# Output:
# Forwarding https://abc123.ngrok.io -> http://localhost:8000

Save the HTTPS URL for ChatGPT connector setup.

MCP Inspector

Test MCP servers without ChatGPT integration:

Installation

npx @modelcontextprotocol/inspector@latest

Usage

# Test local server
npx @modelcontextprotocol/inspector http://localhost:8000/mcp

# Test ngrok URL
npx @modelcontextprotocol/inspector https://abc123.ngrok.io/mcp

Inspector Features

• List Tools - View all available tools and schemas
• Call Tools - Execute tools with test parameters
• View Resources - List and fetch MCP resources
• Test Widgets - Preview widget HTML rendering
• Debug Responses - Inspect structuredContent and _meta

ChatGPT Connector Setup

Enable Developer Mode

1. Open ChatGPT settings
2. Navigate to "Developer settings"
3. Enable "Developer mode"

Create Connector

1. Go to ChatGPT → Settings → Connectors
2. Click "Add connector"
3. Enter your MCP server URL (HTTPS required)
4. Name your connector
5. Save

Add to Conversation

1. Start a new conversation
2. Click the connector icon
3. Select your connector
4. Test with natural language prompts

Testing Workflow

1. Local Server Test

# Start server
python server.py

# Test with curl
curl http://localhost:8000/mcp \
  -X POST \
  -H "Content-Type: application/json" \
  -d '{"method": "tools/list"}'

2. MCP Inspector Test

# Run inspector
npx @modelcontextprotocol/inspector http://localhost:8000/mcp

# In inspector:
# - Verify tools list correctly
# - Test each tool with sample inputs
# - Check widget rendering

3. ngrok Integration Test

# Expose server
ngrok http 8000

# Test HTTPS endpoint
curl https://abc123.ngrok.io/mcp \
  -X POST \
  -H "Content-Type: application/json" \
  -d '{"method": "tools/list"}'

4. ChatGPT Integration Test

1. Add connector with ngrok URL
2. Test natural language tool invocation
3. Verify widget rendering
4. Check response accuracy

Debugging

Enable Debug Logging

Python:

import logging
logging.basicConfig(level=logging.DEBUG)

mcp = FastMCP("server", debug=True)

TypeScript:

const server = new Server(
  { name: "server", version: "1.0.0" },
  { capabilities: { tools: {} }, debug: true }
);

Common Issues

openai/outputTemplate

Request Logging

Python:

@mcp.middleware
async def log_requests(request, call_next):
    print(f"Request: {request.method} - {request.params}")
    response = await call_next(request)
    print(f"Response: {response}")
    return response

Production Deployment

Hosting Options

Railway Deployment

# Install Railway CLI
npm install -g @railway/cli

# Login and init
railway login
railway init

# Deploy
railway up

Docker Deployment

Dockerfile (Python):

FROM python:3.11-slim

WORKDIR /app
COPY requirements.txt .
RUN pip install -r requirements.txt

COPY . .

EXPOSE 8000
CMD ["python", "server.py"]

Build and run:

docker build -t mcp-server .
docker run -p 8000:8000 mcp-server

Environment Variables

Production configuration:

# .env.production
PORT=8000
HOST=0.0.0.0
DATABASE_URL=postgresql://...
API_KEY=your_production_key
LOG_LEVEL=INFO

Health Checks

Add a health check endpoint:

Python:

@mcp.tool()
def health_check() -> dict:
    """Server health check."""
    return {
        "structuredContent": {
            "status": "healthy",
            "version": "1.0.0",
            "timestamp": datetime.now().isoformat()
        }
    }

Production Checklist

• HTTPS enabled (required)
• Environment variables configured
• Debug mode disabled
• Logging configured
• Error handling implemented
• Rate limiting enabled
• Health check endpoint
• Monitoring/alerting setup
• Database backups (if applicable)
• CORS configured correctly

Troubleshooting

Server Not Responding

# Check if server is running
curl http://localhost:8000/mcp

# Check port binding
lsof -i :8000

# Check logs
tail -f server.log

Widget Not Loading

1. Check resource registration
2. Verify

   mimeType: "text/html+skybridge"

3. Inspect

   _meta.openai/outputTemplate

4. Test widget HTML in MCP Inspector

Tool Call Failures

1. Verify tool name matches exactly
2. Check inputSchema validation
3. Review error response format
4. Test with MCP Inspector first

Additional Resources

Reference Files

For detailed deployment guides:

• references/hosting-guide.md

  - Platform-specific deployment

• references/monitoring.md

  - Logging and monitoring setup

Example Files

Working examples in

examples/

• examples/Dockerfile

  - Docker deployment example

• examples/railway.json

  - Railway configuration

Official Documentation

• Apps SDK Deployment: https://developers.openai.com/apps-sdk/build/deploy-your-app/
• ChatGPT Connectors: https://help.openai.com/en/articles/12584461-developer-mode-apps-and-full-mcp-connectors-in-chatgpt-beta