Database Marketplace (OpenAI Apps SDK sample)

This project demonstrates how to build an OpenAI Apps SDK integration that lets an agent sell pre-hosted sandbox databases. The assistant can browse the catalog, create lightweight purchase orders, and run read-only SQL queries against curated datasets via an MCP server that also renders an inline marketplace component.

Features

• Catalog discovery – list available database plans with pricing, schema context, and sample queries.
• Purchase workflow – capture structured purchase intents and return order receipts for the hosting team.
• Hosted SQL execution – run read-only SQL against synthetic datasets (CRM, e-commerce, IoT telemetry) so the agent can answer product questions with live results.
• MCP-first – exposes the catalog, ordering, and SQL tools through the official @modelcontextprotocol/sdk server plus a custom HTML component.

Project structure

├── app.yaml               # Apps SDK manifest pointing at the MCP server
├── assets/database-marketplace.html  # Inline widget rendered by ChatGPT
├── config/catalog.js      # Database plan definitions, schema, and seed data
├── src/
│   ├── server.js          # MCP SSE server wiring tools and resources
│   └── services/
│       ├── databaseService.js  # Initializes sql.js engines per plan
│       └── orderService.js     # In-memory order receipts
└── package.json

Getting started

1. Install dependencies

   npm install

   Use Node.js 20.11+ (for the --env-file flag) or Deno with Node compatibility. The project relies on sql.js which bundles SQLite to run in memory, so no native compilation is required.

2. Run the MCP server locally

   npm run dev

   The SSE server boots on http://localhost:8000 by default and seeds each plan with synthetic data.

3. Verify with MCP Inspector

   • Start MCP Inspector and connect it to http://localhost:8000/mcp with message posts proxied to http://localhost:8000/mcp/messages.
   • List tools and resources to confirm the database marketplace widget renders with structured content.

4. Expose the server to the Apps SDK

   • Update app.yaml with your deployed HTTPS base URL and contact details.
   • Publish the manifest and SSE endpoints following the Apps SDK deployment guide.
   • Create an App in the OpenAI UI, point it at app.yaml, and allow the assistant to call the MCP tools.

5. Playbook for agents

   • Call list_database_plans to summarize available datasets and pricing.
   • Use describe_database_plan to fetch schema context and sample queries for a specific plan.
   • When the user is ready, invoke purchase_database_plan with buyer details to log the purchase intent.
   • For ad-hoc analytics, call run_database_query with a read-only SELECT statement and summarize the results back to the user.

Security notes

• SQL execution is restricted to single SELECT statements. Attempts to run mutating commands return validation errors.
• Order receipts are held in memory only; plug in your billing or CRM system before using in production.
• Add authentication (API keys or OAuth) before exposing the service publicly.

Next steps

• Persist orders in a database and emit webhooks for downstream invoicing.
• Attach per-customer credentials so each buyer gets isolated compute and storage.
• Layer usage metering and alerting for long-running SQL jobs.

License

MIT