Openai Apps Sdk Builder

OpenAI Apps SDK Builder Skill

Overview

This skill guides Claude in creating OpenAI Apps SDK applications - interactive apps that run inside ChatGPT using the Model Context Protocol (MCP). These apps combine conversational AI with rich UI components (widgets) that appear inline in chat.

When to Use This Skill

Use this skill when users request:

• "Create an OpenAI app for [use case]"
• "Build a ChatGPT app that [does something]"
• "Make an MCP server for [functionality]"
• Interactive tools/widgets for ChatGPT (maps, galleries, players, forms, etc.)

Core Concepts

What is an OpenAI Apps SDK App?

An OpenAI Apps SDK app consists of three integrated components:

1. MCP Server: Backend that exposes tools (functions) ChatGPT can call
2. UI Components/Widgets: React-based interactive interfaces rendered in ChatGPT
3. Metadata Bridge: The _meta.openai/outputTemplate that connects tools to UI

How It Works

User Query → ChatGPT decides to use tool → MCP Server executes tool 
→ Returns structured data + metadata → ChatGPT renders widget with data

Architecture Patterns

MCP Server (Backend)

The server implements three key capabilities:

1. List Tools - Advertises available tools with JSON schemas
2. Call Tools - Executes tool logic and returns structured content
3. Return Widgets - Includes metadata pointing to UI components

Widget (Frontend)

React components that:

• Run in an iframe within ChatGPT
• Communicate via window.openai API
• Receive data through tool outputs
• Can call back to MCP server tools
• Support multiple display modes (inline, fullscreen, PiP)

Implementation Guide

Language/Framework Choices

Python (Recommended for rapid prototyping)

• FastMCP or official Python MCP SDK
• FastAPI for HTTP transport
• Best for data-heavy apps, ML integration, rapid development

TypeScript/Node (Recommended for React-heavy apps)

• Official @modelcontextprotocol/sdk
• Express or native Node HTTP
• Best for complex UI, existing Node infrastructure

Project Structure

my-app/
├── src/                          # Widget source code (React)
│   ├── pizza-map/
│   │   ├── index.tsx
│   │   ├── styles.css
│   │   └── types.ts
│   ├── use-openai-global.ts     # Hooks for window.openai
│   └── use-max-height.ts        # Layout utilities
├── assets/                       # Built widget bundles
│   ├── pizza-map-[hash].html
│   ├── pizza-map-[hash].js
│   └── pizza-map-[hash].css
├── server_python/                # Python MCP server
│   ├── main.py
│   └── requirements.txt
├── server_node/                  # TypeScript MCP server
│   ├── src/
│   │   └── index.ts
│   └── package.json
├── build-all.mts                 # Vite build orchestrator
├── vite.config.ts
└── package.json

Python MCP Server Implementation

Setup (FastAPI + FastMCP)

# requirements.txt
fastapi
uvicorn
fastmcp
pydantic

# main.py
from fastapi import FastAPI
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import FileResponse, Response
from fastmcp import FastMCP
import json

app = FastAPI()
mcp = FastMCP(name="My App Server")

# CORS for ChatGPT
app.add_middleware(
    CORSMiddleware,
    allow_origins=["https://chatgpt.com"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# Serve widget HTML
@app.get("/components/{component_name}.html")
async def serve_component(component_name: str):
    return FileResponse(
        f"assets/{component_name}.html",
        media_type="text/html+skybridge"  # CRITICAL: Use this MIME type
    )

# MCP endpoint (Streamable HTTP)
@app.post("/mcp")
async def handle_mcp(request: Request):
    # Handle MCP protocol - use FastMCP or manual implementation
    return mcp.handle_request(await request.json())

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8000)

Registering Resources (UI Templates)

# Register a UI resource that the server can serve
@mcp.resource("pizza-map")
async def get_pizza_map_resource():
    """Returns the HTML template for the pizza map widget"""
    return {
        "contents": [{
            "uri": "ui://widget/pizza-map.html",
            "mimeType": "text/html+skybridge",  # CRITICAL MIME type
            "text": """
                <div id="pizzaz-root"></div>
                <link rel="stylesheet" href="https://your-cdn.com/pizza-map-abc123.css">
                <script type="module" src="https://your-cdn.com/pizza-map-abc123.js"></script>
            """.strip(),
            "_meta": {
                "openai/widgetDescription": "Interactive map showing pizza locations with ratings",
                "openai/widgetPrefersBorder": True,
                "openai/widgetCSP": {
                    "connect_domains": [],
                    "resource_domains": ["https://your-cdn.com", "https://api.mapbox.com"]
                },
                "openai/widgetDomain": "https://chatgpt.com"
            }
        }]
    }

Registering Tools

@mcp.tool()
async def find_pizza_places(location: str, max_results: int = 10) -> dict:
    """
    Find pizza restaurants near a location.
    
    Args:
        location: City or address to search near
        max_results: Maximum number of results to return
    
    Returns:
        Dictionary with pizza places and map data
    """
    # Your business logic here
    places = [
        {
            "id": "1",
            "name": "Pizza Palace",
            "rating": 4.5,
            "coords": [-73.935242, 40.730610],
            "description": "Classic New York style pizza"
        },
        # ... more places
    ]
    
    return {
        # Plain text for conversation transcript
        "content": [{
            "type": "text",
            "text": f"Found {len(places)} pizza places near {location}"
        }],
        
        # Structured data that ChatGPT can reason about
        "structuredContent": {
            "places": places,
            "location": location
        },
        
        # Metadata for widget rendering (NOT shown to model)
        "_meta": {
            "openai/outputTemplate": "ui://widget/pizza-map.html",
            "openai/toolInvocation/invoking": "Searching for pizza...",
            "openai/toolInvocation/invoked": "Found pizza places",
            
            # Additional data for widget only (not shown to model)
            "mapSettings": {
                "center": [-73.935242, 40.730610],
                "zoom": 12
            }
        }
    }

Tool Metadata Annotations

# Read-only hint (for tools that don't modify state)
@mcp.tool(
    _meta={
        "openai/readOnlyHint": True
    }
)
async def list_favorites() -> dict:
    """List user's favorite pizza places without modifying anything"""
    pass

# Custom invocation messages
@mcp.tool(
    _meta={
        "openai/toolInvocation/invoking": "Calculating route...",
        "openai/toolInvocation/invoked": "Route calculated"
    }
)
async def get_directions(from_loc: str, to_loc: str) -> dict:
    """Get directions between two locations"""
    pass

# Locale-aware tools
@mcp.tool()
async def recommend_cafe(location: str, context: dict) -> dict:
    """Recommend a cafe based on user location"""
    locale = context.get("_meta", {}).get("openai/locale", "en")
    user_location = context.get("_meta", {}).get("openai/userLocation", {})
    
    # Use locale and location hints (never for auth!)
    return {...}

TypeScript/Node MCP Server Implementation

Setup

// package.json
{
  "dependencies": {
    "@modelcontextprotocol/sdk": "^1.0.0",
    "express": "^4.18.0",
    "zod": "^3.22.0"
  }
}

// src/index.ts
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { z } from "zod";
import express from "express";

const server = new McpServer({
  name: "My App Server",
  version: "1.0.0"
});

const app = express();

// CORS middleware
app.use((req, res, next) => {
  res.header("Access-Control-Allow-Origin", "https://chatgpt.com");
  res.header("Access-Control-Allow-Methods", "GET, POST, OPTIONS");
  res.header("Access-Control-Allow-Headers", "Content-Type");
  next();
});

// Serve widget bundles
app.get("/components/:name.html", (req, res) => {
  res.setHeader("Content-Type", "text/html+skybridge");
  res.sendFile(`assets/${req.params.name}.html`);
});

app.listen(8000);

Registering Resources

server.registerResource(
  "pizza-map",
  "ui://widget/pizza-map.html",
  {},
  async () => ({
    contents: [{
      uri: "ui://widget/pizza-map.html",
      mimeType: "text/html+skybridge",
      text: `
        <div id="pizzaz-root"></div>
        <link rel="stylesheet" href="https://your-cdn.com/pizza-map-abc123.css">
        <script type="module" src="https://your-cdn.com/pizza-map-abc123.js"></script>
      `.trim(),
      _meta: {
        "openai/widgetDescription": "Interactive map showing pizza locations",
        "openai/widgetPrefersBorder": true
      }
    }]
  })
);

Registering Tools

server.registerTool(
  "find_pizza_places",
  {
    title: "Find Pizza Places",
    description: "Search for pizza restaurants near a location",
    inputSchema: z.object({
      location: z.string(),
      maxResults: z.number().int().min(1).max(50).optional()
    }),
    _meta: {
      "openai/outputTemplate": "ui://widget/pizza-map.html",
      "openai/readOnlyHint": true
    }
  },
  async ({ location, maxResults = 10 }) => {
    const places = await searchPizzaPlaces(location, maxResults);
    
    return {
      content: [{
        type: "text",
        text: `Found ${places.length} pizza places near ${location}`
      }],
      structuredContent: { places, location },
      _meta: {
        mapSettings: {
          center: places[0]?.coords || [0, 0],
          zoom: 12
        }
      }
    };
  }
);

Widget Development

React Component Structure

// src/pizza-map/index.tsx
import React, { useEffect, useRef } from "react";
import { createRoot } from "react-dom/client";
import mapboxgl from "mapbox-gl";
import "mapbox-gl/dist/mapbox-gl.css";
import { useOpenAiGlobal } from "../use-openai-global";
import { useMaxHeight } from "../use-max-height";

function PizzaMap() {
  // Access tool input/output from window.openai
  const toolInput = useOpenAiGlobal("toolInput");
  const toolOutput = useOpenAiGlobal("toolOutput");
  const metadata = useOpenAiGlobal("toolResponseMetadata");
  
  // Access theme and display mode
  const theme = useOpenAiGlobal("theme");
  const displayMode = useOpenAiGlobal("displayMode");
  
  // Handle responsive height
  const maxHeight = useMaxHeight();
  
  const mapRef = useRef<mapboxgl.Map | null>(null);
  
  useEffect(() => {
    // Initialize map
    const map = new mapboxgl.Map({
      container: "map",
      style: theme === "dark" 
        ? "mapbox://styles/mapbox/dark-v11"
        : "mapbox://styles/mapbox/light-v11",
      center: metadata?.mapSettings?.center || [0, 0],
      zoom: metadata?.mapSettings?.zoom || 2
    });
    
    mapRef.current = map;
    
    // Add markers for each place
    toolOutput?.places?.forEach(place => {
      new mapboxgl.Marker()
        .setLngLat(place.coords)
        .setPopup(
          new mapboxgl.Popup().setHTML(
            `<h3>${place.name}</h3><p>Rating: ${place.rating}</p>`
          )
        )
        .addTo(map);
    });
    
    return () => map.remove();
  }, [toolOutput, theme, metadata]);
  
  // Handle display mode changes
  useEffect(() => {
    if (displayMode === "fullscreen") {
      // Expand to full capabilities
      mapRef.current?.resize();
    }
  }, [displayMode]);
  
  return (
    <div style={{ height: maxHeight, width: "100%" }}>
      <div id="map" style={{ height: "100%", width: "100%" }} />
    </div>
  );
}

// Mount the component
const root = document.getElementById("pizzaz-root");
if (root) {
  createRoot(root).render(<PizzaMap />);
}

Essential Hooks

// use-openai-global.ts
import { useEffect, useState } from "react";

export function useOpenAiGlobal<K extends keyof OpenAiGlobals>(
  key: K
): OpenAiGlobals[K] | undefined {
  const [value, setValue] = useState<OpenAiGlobals[K]>();
  
  useEffect(() => {
    // Get initial value
    if (window.openai?.[key]) {
      setValue(window.openai[key]);
    }
    
    // Listen for changes
    const handler = (event: SetGlobalsEvent) => {
      if (key in event.detail.globals) {
        setValue(event.detail.globals[key]);
      }
    };
    
    window.addEventListener("openai:set_globals", handler);
    return () => window.removeEventListener("openai:set_globals", handler);
  }, [key]);
  
  return value;
}

// Specialized hooks
export function useToolInput() {
  return useOpenAiGlobal("toolInput");
}

export function useToolOutput() {
  return useOpenAiGlobal("toolOutput");
}

export function useTheme() {
  return useOpenAiGlobal("theme");
}

export function useDisplayMode() {
  return useOpenAiGlobal("displayMode");
}

Widget State Management

// For persisting state across sessions
import { useWidgetState } from "../use-widget-state";

function MyWidget() {
  const [state, setState] = useWidgetState({
    favorites: [],
    selectedId: null
  });
  
  const addFavorite = (id: string) => {
    setState(prev => ({
      ...prev,
      favorites: [...prev.favorites, id]
    }));
  };
  
  // Note: Widget state is exposed to ChatGPT model
  // Keep it small (<4k tokens) for performance
  
  return <div>...</div>;
}

Calling MCP Tools from Widget

function InteractiveWidget() {
  const [loading, setLoading] = useState(false);
  
  const handleAction = async () => {
    setLoading(true);
    
    try {
      const result = await window.openai.callTool("update_preference", {
        preference: "dark_mode",
        value: true
      });
      
      console.log("Tool result:", result);
    } catch (error) {
      console.error("Tool call failed:", error);
    } finally {
      setLoading(false);
    }
  };
  
  return <button Preference</button>;
}

Display Mode Management

function ExpandableWidget() {
  const displayMode = useDisplayMode();
  
  const requestFullscreen = () => {
    window.openai.requestDisplayMode({ mode: "fullscreen" });
  };
  
  const requestPiP = () => {
    window.openai.requestDisplayMode({ mode: "pip" });
  };
  
  return (
    <div>
      {displayMode !== "fullscreen" && (
        <button
          Expand to Fullscreen
        </button>
      )}
      
      {displayMode === "inline" && (
        <button
          Pop Out (PiP)
        </button>
      )}
    </div>
  );
}

Building and Bundling Widgets

Vite Configuration

// vite.config.ts
import { defineConfig } from "vite";
import react from "@vitejs/plugin-react";
import { resolve } from "path";
import { readdirSync } from "fs";

// Auto-discover all widget entry points
const widgets = readdirSync("src").filter(dir => 
  !dir.startsWith("use-") && !dir.includes(".")
);

const input = Object.fromEntries(
  widgets.map(widget => [widget, resolve(__dirname, `src/${widget}/index.tsx`)])
);

export default defineConfig({
  plugins: [react()],
  build: {
    outDir: "assets",
    rollupOptions: {
      input,
      output: {
        entryFileNames: "[name]-[hash].js",
        chunkFileNames: "[name]-[hash].js",
        assetFileNames: "[name]-[hash].[ext]"
      }
    }
  }
});

Build Script

// build-all.mts
import { build } from "vite";
import { writeFileSync } from "fs";

async function buildAll() {
  // Build all widgets
  await build();
  
  // Generate HTML wrappers with hashed asset references
  const widgets = ["pizza-map", "pizza-carousel", "pizza-video"];
  
  for (const widget of widgets) {
    const hash = "abc123"; // Extract from build output
    const html = `
      <!DOCTYPE html>
      <html>
        <head>
          <meta charset="UTF-8">
          <meta name="viewport" content="width=device-width, initial-scale=1.0">
          <link rel="stylesheet" href="https://your-cdn.com/${widget}-${hash}.css">
        </head>
        <body>
          <div id="pizzaz-root"></div>
          <script type="module" src="https://your-cdn.com/${widget}-${hash}.js"></script>
        </body>
      </html>
    `.trim();
    
    writeFileSync(`assets/${widget}-${hash}.html`, html);
  }
}

buildAll();

Best Practices

1. Tool Design

DO:

• Use clear, descriptive tool names (e.g., find_pizza_places not search)
• Write detailed docstrings - they become tool descriptions
• Keep tool focused on one clear action
• Mark read-only tools with openai/readOnlyHint
• Return both human-readable text AND structured data

DON'T:

• Use generic names like process or handle
• Create tools that do multiple unrelated things
• Rely on _meta fields for authorization
• Expose sensitive data in structuredContent

2. Widget Development

DO:

• Use text/html+skybridge MIME type for all widget resources
• Handle theme changes (light/dark mode)
• Support all display modes (inline, fullscreen, PiP)
• Keep widget state small (<4k tokens)
• Use semantic HTML and WCAG AA contrast ratios
• Test on mobile, tablet, and desktop

DON'T:

• Use HTML <form> elements (blocked in iframe)
• Store sensitive data in widget state (it's visible to the model)
• Assume localStorage/sessionStorage work (they don't in iframe)
• Hard-code aspect ratios that might distort
• Include your logo (ChatGPT adds it automatically)

3. Data Flow

User Input
    ↓
ChatGPT (decides to use tool)
    ↓
MCP Server Tool Execution
    ↓
Returns:
  - content: Text for conversation (visible)
  - structuredContent: Data for model reasoning (visible)
  - _meta: Widget data + config (NOT visible to model)
    ↓
ChatGPT renders widget using _meta.openai/outputTemplate
    ↓
Widget receives data via window.openai
    ↓
Widget can call tools back via window.openai.callTool()

4. Security

DO:

• Use CORS to restrict origins to https://chatgpt.com
• Validate all tool inputs with schemas
• Implement OAuth 2.1 for user authentication
• Use CSP (Content Security Policy) in widget metadata
• Rate limit tool calls
• Validate webhook signatures

DON'T:

• Trust _meta["openai/userLocation"] for authorization
• Store secrets in widget code or metadata
• Allow unrestricted CORS origins
• Skip input validation

5. Performance

DO:

• Keep widget bundles small (<500KB ideally)
• Use code splitting for large widgets
• Optimize images (WebP, compression)
• Cache static assets aggressively
• Use CDN for widget bundles
• Minimize widget state updates

DON'T:

• Load entire libraries for small features
• Make excessive API calls from widgets
• Store large datasets in widget state
• Trigger unnecessary re-renders

Testing

Local Development

# Terminal 1: Build widgets
npm run build

# Terminal 2: Serve static assets with CORS
npm run serve  # Assets at http://localhost:4444

# Terminal 3: Run MCP server
# Python
uvicorn main:app --reload --port 8000

# Node
npm run dev

Testing with ngrok

# Expose local server to internet
ngrok http 8000

# Use ngrok URL in ChatGPT Developer Mode
# Settings > Connectors > Add Connector
# MCP URL: https://abc123.ngrok-free.app/mcp

MCP Inspector

# Test tools without ChatGPT
npx @modelcontextprotocol/inspector http://localhost:8000/mcp

# Validates:
# - Tool schemas
# - Resource responses
# - Metadata structure
# - Widget rendering

Deployment

Requirements

• HTTPS endpoint (required for production)
• Low cold-start latency (<2s ideal)
• Support for streaming HTTP or SSE transport
• CORS configured for https://chatgpt.com

Hosting Options

Serverless (AWS Lambda, Google Cloud Functions)

• Pros: Auto-scaling, pay-per-use
• Cons: Cold starts, 15-30s timeout limits
• Best for: Low-traffic apps, bursty workloads

Container (Cloud Run, Fargate, Railway)

• Pros: Consistent performance, longer timeouts
• Cons: More complex setup, always-on costs
• Best for: Production apps, steady traffic

VPS (DigitalOcean, Linode, Hetzner)

• Pros: Full control, predictable costs
• Cons: Manual scaling, server management
• Best for: High-performance needs, custom infrastructure

Environment Variables

# Production
OPENAI_API_KEY=sk-...              # If calling OpenAI APIs
WIDGET_CDN_URL=https://cdn.example.com
DATABASE_URL=postgresql://...
REDIS_URL=redis://...
OAUTH_CLIENT_ID=...
OAUTH_CLIENT_SECRET=...

# Development
WIDGET_CDN_URL=http://localhost:4444
DEBUG=true

CDN Setup for Widgets

# Build with production CDN URL
export WIDGET_CDN_URL=https://persistent.oaistatic.com/your-app

npm run build

# Upload assets/ to CDN
aws s3 sync assets/ s3://your-bucket/
# or
gsutil -m rsync -r assets/ gs://your-bucket/

Common Widget Patterns

1. Map Widget

• Libraries: Mapbox GL, Leaflet
• Use cases: Restaurant finder, store locator, real estate
• Features: Markers, popups, clustering, directions

2. Carousel/Gallery

• Libraries: Embla, Swiper
• Use cases: Product browsing, image galleries, portfolios
• Features: Touch gestures, lazy loading, thumbnails

3. Video/Audio Player

• Libraries: Video.js, Plyr
• Use cases: Courses, music, podcasts
• Features: Timeline, chapters, playback controls, captions

4. List/Grid

• Libraries: React Virtualized, TanStack Virtual
• Use cases: Search results, catalogs, feeds
• Features: Infinite scroll, filtering, sorting, actions

5. Form/Survey

• Libraries: React Hook Form, Formik
• Use cases: Booking, checkout, data collection
• Features: Validation, multi-step, autosave

6. Chart/Visualization

• Libraries: Recharts, D3, Chart.js
• Use cases: Analytics, reports, dashboards
• Features: Interactive legends, tooltips, drill-down

Troubleshooting

Widget Not Rendering

Symptom: ChatGPT shows text response but no widget

Possible Causes:

1. Wrong MIME type - MUST be text/html+skybridge
2. Missing _meta["openai/outputTemplate"] in tool response
3. Resource URI doesn't match template reference
4. CORS not allowing https://chatgpt.com
5. Widget bundle has JavaScript errors

Fix:

# Ensure correct MIME type
return FileResponse(
    "widget.html",
    media_type="text/html+skybridge"  # ← Critical!
)

# Ensure metadata is correct
return {
    "_meta": {
        "openai/outputTemplate": "ui://widget/my-widget.html"  # ← Must match resource URI
    }
}

Tool Not Being Called

Symptom: ChatGPT responds with general knowledge instead of using your tool

Possible Causes:

1. Tool name too generic
2. Docstring/description unclear
3. Tool not registered properly
4. MCP server not responding

Fix:

# BAD: Generic name and description
@mcp.tool()
async def search(query: str):
    """Search for results"""
    pass

# GOOD: Specific name and clear description
@mcp.tool()
async def find_pizza_restaurants(location: str, max_results: int = 10):
    """
    Search for pizza restaurants near a specific location.
    
    Use this tool when the user wants to find pizza places,
    pizzerias, or Italian restaurants that serve pizza.
    """
    pass

CORS Errors

Symptom: Browser console shows CORS error when loading widget

Fix:

# Python/FastAPI
from fastapi.middleware.cors import CORSMiddleware

app.add_middleware(
    CORSMiddleware,
    allow_origins=["https://chatgpt.com"],  # Be specific!
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

Widget State Not Persisting

Symptom: Widget state resets on refresh or new tool calls

Cause: Not using setWidgetState properly

Fix:

// Use the official hook
const [state, setState] = useWidgetState({ count: 0 });

// Updates persist across sessions
setState(prev => ({ count: prev.count + 1 }));

Example Apps to Reference

When building, reference these official examples from the OpenAI repository:

1. Pizzaz List - Card list with favorites and CTAs
2. Pizzaz Carousel - Horizontal scroller for media
3. Pizzaz Map - Mapbox integration with fullscreen
4. Pizzaz Album - Stacked gallery view
5. Pizzaz Video - Video player with overlays
6. Solar System - 3D interactive visualization

Checklist Before Publishing

• Tool names are specific and descriptive
• All tools have clear docstrings
• Read-only tools marked with openai/readOnlyHint
• CORS restricted to https://chatgpt.com
• Widget uses text/html+skybridge MIME type
• Widget handles light and dark themes
• Widget supports all display modes (inline, fullscreen, PiP)
• Widget is accessible (WCAG AA contrast, alt text, keyboard nav)
• No HTML forms used (use divs + onClick instead)
• Widget state is minimal (<4k tokens)
• Tested with MCP Inspector
• Tested in ChatGPT Developer Mode
• HTTPS endpoint configured for production
• OAuth implemented if needed
• Rate limiting configured
• Error handling for all tool calls
• Widget bundles optimized and hosted on CDN

Resources

• Official Docs: https://developers.openai.com/apps-sdk/
• Examples Repo: https://github.com/openai/openai-apps-sdk-examples
• MCP Spec: https://spec.modelcontextprotocol.io/
• Python SDK: https://github.com/modelcontextprotocol/python-sdk
• TypeScript SDK: https://github.com/modelcontextprotocol/typescript-sdk
• FastMCP: https://github.com/jlowin/fastmcp

Summary

When building OpenAI Apps SDK applications:

1. Start with the MCP server - Define tools with clear schemas
2. Build widget components - React apps using window.openai API
3. Connect with metadata - Use _meta["openai/outputTemplate"]
4. Test locally - Use MCP Inspector and ngrok
5. Deploy with HTTPS - Required for ChatGPT integration
6. Follow design guidelines - Accessible, responsive, theme-aware

The key insight: Apps SDK combines conversational AI with rich UI by using MCP as the protocol bridge. The model decides when to call tools, tools return data + UI metadata, and widgets render the interactive experience - all within the ChatGPT chat interface.