ChatGPT Work runs on MCP. Build for the standard, not the app.

The headline today is ChatGPT Work, OpenAI's new agent that builds finished documents and web apps on its own. The part worth your attention outlasts the headline: what it plugs into. ChatGPT reaches your business systems through MCP connectors — the Model Context Protocol — and that layer, not the shiny agent on top of it, is the thing to design around. Because MCP isn't OpenAI's. It's a shared standard, and building to it is how you avoid rewriting the same integration for every AI vendor that comes knocking.

What actually happened

To do real work, ChatGPT has to see your data. OpenAI's approach is a set of partner-built, OpenAI-reviewed MCP access connectors — its help documentation lists integrations including Stripe, Vercel, Monday.com, Amplitude, Hex, Egnyte, and more, available to Enterprise, Edu, Business, Pro, and Plus workspaces with connectors enabled. These aren't bespoke one-offs. Vercel's MCP server, for example, has supported ChatGPT since 2025, letting an agent reach protected deployments and analyze build logs through a single OAuth-secured endpoint.

That's the pattern. A vendor stands up one MCP server describing what an agent can read and do; any MCP-speaking client connects to it. The Model Context Protocol was published by Anthropic as an open standard and has since been adopted across the major agent platforms — we've written before about MCP servers being scoped read-only by design. Today's ChatGPT Work launch is the loudest proof yet that MCP is becoming the default way agents touch your systems.

Why it matters for your business

Here's the operator's angle. If you build a custom integration so ChatGPT can read your orders, and next quarter your team wants Claude or Copilot to do the same, you build it again. And again. That's the trap of integrating with applications — you chase whichever agent is winning this month and pay the tax each time.

Build to the standard instead. Expose your systems — your database, your order history, your internal tools — through one MCP interface you own, with the scopes and governance defined by you: what's readable, what's writable, who's allowed. Then any agent, ChatGPT Work included, connects to the same endpoint. You stop rewriting integrations and start controlling exactly how much of your business an agent can see and touch. The connector is where the leverage is. Own it, define its limits, and let the agents compete for the right to use it.

Sources: OpenAI — Developer mode and MCP apps in ChatGPT, Vercel — ChatGPT is now supported on Vercel MCP.