ChatGPT Apps Builder

Complete workflow for building, testing, and deploying ChatGPT Apps from concept to production.

Commands

• - /chatgpt-apps new - Create a new ChatGPT App
• INLINECODE1 - Add an MCP tool to your app
• INLINECODE2 - Add a widget to your app
• INLINECODE3 - Configure authentication
• INLINECODE4 - Set up database
• INLINECODE5 - Validate your app
• INLINECODE6 - Run tests
• INLINECODE7 - Deploy to production
• INLINECODE8 - Resume working on an app

---

Table of Contents

1. 1. Create New App
2. Add MCP Tool
3. Add Widget
4. Add Authentication
5. Add Database
6. Generate Golden Prompts
7. Validate App
8. Test App
9. Deploy App
10. Resume App

---

1. Create New App

Purpose: Create a new ChatGPT App from concept to working code.

Workflow

Phase 1: Conceptualization

1. 1. Ask for the app idea

"What ChatGPT App would you like to build? Describe what it does and the problem it solves."

1. 2. Analyze against UX Principles

- Conversational Leverage: What can users accomplish through natural language? - Native Fit: How does this integrate with ChatGPT's conversational flow? - Composability: Can tools work independently and combine with other apps?

1. 3. Check for Anti-Patterns

- Static website content display - Complex multi-step workflows requiring external tabs - Duplicating ChatGPT's native capabilities - Ads or upsells

1. 4. Define Use Cases

Create 3-5 primary use cases with user stories.

Phase 2: Design

1. 1. Tool Topology

- Query tools (readOnlyHint: true) - Mutation tools (destructiveHint: false) - Destructive tools (destructiveHint: true) - Widget tools (return UI with _meta) - External API tools (openWorldHint: true)

1. 2. Widget Design

For each widget: - id - unique identifier (kebab-case) - name - display name - description - what it shows - mockData - sample data for preview

1. 3. Data Model

Design entities and relationships.

1. 4. Auth Requirements

- Single-user (no auth needed) - Multi-user (Auth0 or Supabase Auth)

Phase 3: Implementation

Generate complete application with this structure:

CODEBLOCK0

Critical Requirements:

• - Server class from INLINECODE14
• INLINECODE15 for session management
• Widget URIs: INLINECODE16
• Widget MIME type: INLINECODE17
• INLINECODE18 in tool responses
• INLINECODE19 with openai/outputTemplate on tools

Phase 4: Testing

• - Run setup: INLINECODE21
• Start dev: INLINECODE22
• Preview widgets: INLINECODE23
• Test MCP connection

Phase 5: Deployment

• - Generate Dockerfile and render.yaml
• Deploy to Render
• Configure ChatGPT connector

---

2. Add MCP Tool

Purpose: Add a new MCP tool to your ChatGPT App.

Workflow

1. 1. Gather Information

- What does this tool do? - What inputs does it need? - What does it return?

1. 2. Classify Tool Type

- Query (readOnlyHint: true) - Fetches data - Mutation (destructiveHint: false) - Creates/updates data - Destructive (destructiveHint: true) - Deletes data - Widget - Returns UI content - External (openWorldHint: true) - Calls external APIs

1. 3. Design Input Schema

Create Zod schema with appropriate types and descriptions.

1. 4. Generate Tool Handler

Use chatgpt-mcp-generator agent to create: - Tool handler in server/tools/ - Zod schema export - Type exports - Database queries (if needed)

1. 5. Register Tool

Update server/index.ts with metadata: CODEBLOCK1

1. 6. Update State

Add tool to .chatgpt-app/state.json.

Tool Naming

Use kebab-case: list-items, create-task, INLINECODE30

Annotations Guide

Scenario	readOnlyHint	destructiveHint	openWorldHint
List/Get	true	false	false
Create/Update

false | false | false | | Delete | false | true | false | | External API | varies | varies | true |

---

3. Add Widget

Purpose: Add inline HTML widgets with HTML/CSS/JS and Apps SDK integration.

5 Widget Patterns

1. 1. Card Grid - Multiple items in grid
2. Stats Dashboard - Key metrics display
3. Table - Tabular data
4. Bar Chart - Simple visualizations
5. Detail Widget - Single item details

Workflow

1. 1. Gather Information

- Widget purpose and data - Visual design (cards, table, chart, etc.) - Interactivity needs

1. 2. Define Data Shape

Document expected structure with TypeScript interface.

1. 3. Add Widget Config

CODEBLOCK2

1. 4. Add Widget HTML

Generate HTML with: - Preview mode support (window.PREVIEW_DATA) - OpenAI Apps SDK integration (window.openai.toolOutput) - Event listeners (openai:set_globals) - Polling fallback (100ms, 10s timeout)

1. 5. Create/Update Tool

Link tool to widget via widgetId.

1. 6. Test Widget

Preview at /preview/{widget-id} with mock data.

Widget HTML Structure

CODEBLOCK3

---

4. Add Authentication

Purpose: Configure authentication using Auth0 or Supabase Auth.

When to Add

• - Multiple users
• Persistent private data per user
• User-specific API credentials

Providers

Auth0:

• - Enterprise-grade
• OAuth 2.1, PKCE flow
• Social logins (Google, GitHub, etc.)

Supabase Auth:

• - Simpler setup
• Email/password default
• Integrates with Supabase database

Workflow

1. 1. Choose Provider

Ask user preference based on needs.

1. 2. Guide Setup

- Auth0: Create application, configure callback URLs, get credentials - Supabase: Already configured with database setup

1. 3. Generate Auth Code

Use chatgpt-auth-generator agent to create: - Session management middleware - User subject extraction - Token validation

1. 4. Update Server

Add auth middleware to protect routes.

1. 5. Update Environment

CODEBLOCK4

1. 6. Test

Verify login flow and user isolation.

---

5. Add Database

Purpose: Configure PostgreSQL database using Supabase.

When to Add

• - Persistent user data
• Multi-entity relationships
• Query/filter capabilities

Workflow

1. 1. Check Supabase Setup

Verify account and project exist.

1. 2. Gather Credentials

- Project URL - Anon key (public) - Service role key (server-side)

1. 3. Define Entities

For each entity, specify: - Fields and types - Relationships - Indexes

1. 4. Generate Schema

Use chatgpt-database-generator agent to create SQL with: - id (UUID primary key) - user_subject (varchar, indexed) - created_at (timestamptz) - updated_at (timestamptz) - RLS policies for user isolation

1. 5. Setup Connection Pool

CODEBLOCK5

1. 6. Apply Migrations

Run SQL in Supabase dashboard or via migration tool.

Query Pattern

Always filter by user_subject:

CODEBLOCK6

---

6. Generate Golden Prompts

Purpose: Generate test prompts to validate ChatGPT correctly invokes tools.

Why Important

• - Measure precision/recall
• Enable iteration
• Post-launch monitoring

3 Categories

1. 1. Direct Prompts - Explicit tool invocation

- "Show me my task list" - "Create a new task called..."

1. 2. Indirect Prompts - Outcome-based, ChatGPT should infer tool

- "What do I need to do today?" - "Help me organize my work"

1. 3. Negative Prompts - Should NOT trigger tool

- "What is a task?" - "Tell me about project management"

Workflow

1. 1. Analyze Tools

Review each tool's purpose and inputs.

1. 2. Generate Prompts

For each tool, create: - 5+ direct prompts - 5+ indirect prompts - 3+ negative prompts - 2+ edge case prompts

1. 3. Best Practices

- Tool descriptions start with "Use this when..." - State limitations clearly - Include examples in descriptions

1. 4. Save Output

Write to .chatgpt-app/golden-prompts.json:

   {
     "toolName": {
       "direct": ["prompt1", "prompt2"],
       "indirect": ["prompt1", "prompt2"],
       "negative": ["prompt1", "prompt2"],
       "edge": ["prompt1", "prompt2"]
     }
   }
   

---

7. Validate App

Purpose: Validation suite before deployment.

10 Validation Checks

1. 1. Required Files

- package.json - tsconfig.server.json - setup.sh (executable) - START.sh (executable) - server/index.ts - .env.example

1. 2. Server Implementation

- Uses Server from MCP SDK - Has StreamableHTTPServerTransport - Session management with Map - Correct request handlers

1. 3. Widget Configuration

- widgets array exists - Each has id, name, description, templateUri, mockData - URIs match pattern INLINECODE47

1. 4. Tool Response Format

- Returns structuredContent (not just content) - Widget tools have _meta with INLINECODE51

1. 5. Resource Handler Format

- MIME type: text/html+skybridge - Returns _meta with serialization and CSP

1. 6. Widget HTML Structure

- Preview mode support - Event listeners for Apps SDK - Polling fallback - Render guard

1. 7. Endpoint Existence

- /health - Health check - /preview - Widget index - /preview/:widgetId - Widget preview - /mcp - MCP endpoint

1. 8. Package.json Scripts

- Has build:server - Has start with HTTP_MODE=true - Has dev with watch mode - NO web build scripts (web/, ui/, client/)

1. 9. Annotation Validation

- readOnlyHint set correctly - destructiveHint for delete operations - openWorldHint for external APIs

1. 10. Database Validation (if enabled)

- Tables have required fields - user_subject indexed - RLS policies enabled

Common Errors

Error	Fix
Missing structuredContent	Add to tool response
Wrong widget URI

Use ui://widget/{id}.html | | No session management | Add Map | | Missing _meta | Add to tool definition and response | | Wrong MIME type | Use text/html+skybridge |

Critical: Check file existence FIRST before other validations!

---

8. Test App

Purpose: Run automated tests using MCP Inspector and golden prompts.

4 Test Categories

1. 1. MCP Protocol

- Server starts without errors - Handles initialize - Lists tools correctly - Lists resources correctly

1. 2. Schema Validation

- Tool schemas are valid Zod - Required fields marked - Types match implementation

1. 3. Widget Tests

- All widgets render in preview mode - Mock data loads correctly - No console errors

1. 4. Golden Prompt Tests

- Direct prompts trigger correct tools - Indirect prompts work as expected - Negative prompts don't trigger tools

Workflow

1. 1. Start Server in Test Mode

CODEBLOCK8

1. 2. Run MCP Inspector

Test protocol compliance: - Initialize connection - List tools - Call each tool with valid inputs - Check responses

1. 3. Schema Validation

Verify schemas compile and match implementation.

1. 4. Golden Prompt Tests

Use ChatGPT to test prompts: - Record which tool was called - Compare to expected tool - Calculate precision/recall

1. 5. Generate Report

CODEBLOCK9

Fixing Failures

For each failure, explain:

• - What failed
• Why it failed
• How to fix (with code example)

---

9. Deploy App

Purpose: Deploy ChatGPT App to Render with PostgreSQL and health checks.

Prerequisites

• - ✅ Validation passed
• ✅ Tests passed
• ✅ Git repository clean
• ✅ Environment variables ready

Workflow

1. 1. Pre-flight Check

- Run validation - Run tests - Check database connection (if enabled)

1. 2. Generate render.yaml

CODEBLOCK10

1. 3. Generate Dockerfile

CODEBLOCK11

1. 4. Deploy

Option A: Automated (if Render MCP available) Use Render MCP agent to deploy.

Option B: Manual
- Push to GitHub
- Connect repo in Render dashboard
- Set environment variables
- Deploy

1. 5. Verify Deployment

- Health check: https://{app}.onrender.com/health - MCP endpoint: https://{app}.onrender.com/mcp - Tool discovery works - Widgets render

1. 6. Configure ChatGPT Connector

- URL: https://{app}.onrender.com/mcp - Test in ChatGPT

---

10. Resume App

Purpose: Resume building an in-progress ChatGPT App.

Workflow

1. 1. Load State

Read .chatgpt-app/state.json: CODEBLOCK12

1. 2. Display Progress

Show current status: - App name - Current phase - Completed items (tools, widgets) - Pending items (auth, validation, deployment)

1. 3. Offer Next Steps

Based on phase:

Concept Phase:
- "Let's design the tools and widgets"
- "Shall we start implementation?"

Implementation Phase:
- "Add another tool?"
- "Add a widget?"
- "Set up authentication?"
- "Set up database?"

Testing Phase:
- "Generate golden prompts?"
- "Run validation?"
- "Run tests?"

Deployment Phase:
- "Deploy to Render?"
- "Configure ChatGPT connector?"

1. 4. Continue Work

Based on user's choice, invoke the appropriate workflow section.

---

Best Practices

1. 1. Always save state after each major step
2. Validate before moving forward (especially before deployment)
3. Use agents for code generation (chatgpt-mcp-generator, chatgpt-auth-generator, etc.)
4. Test at every phase (preview widgets, test tools, run golden prompts)
5. Keep it conversational - guide the user naturally through the workflow
6. Explain trade-offs when offering choices (Auth0 vs Supabase, etc.)
7. Show examples when introducing new concepts

---

State Management

The .chatgpt-app/state.json file tracks progress:

CODEBLOCK13

---

Command Reference

CODEBLOCK14

---

Getting Started

When the user invokes any chatgpt-app command:

1. 1. Check if .chatgpt-app/state.json exists
2. If yes → use Resume App workflow
3. If no → use Create New App workflow

Always guide users through the natural progression:
Concept → Implementation → Testing → Deployment