ChatGPT MCP: How to Connect an MCP Server to ChatGPT

ChatGPT MCP support is real and shipped: since late 2025, ChatGPT can connect to custom MCP servers through connectors in developer mode, letting the model discover and call your server’s tools — including write actions — from inside a conversation. MCP (the Model Context Protocol, introduced by Anthropic and adopted industry-wide) is the same standard Claude and Gemini clients speak, which is exactly the point: one server, every assistant.

OpenAI’s adoption arrived in layers:

This guide covers the generic setup — any MCP server into ChatGPT. If what you want to connect is your own business data (analytics, CRM, revenue) rather than a server you’re coding, the product-specific walkthrough is connecting Claude, GPT and Gemini to your workspace.

Three prerequisites, and the first one filters out most surprises:

• A remote MCP server. ChatGPT reaches your server over the network — there is no local stdio option like desktop clients have. The server should speak MCP’s Streamable HTTP transport over HTTPS, either open or protected by OAuth. If your server currently runs only as a local process, our remote vs local MCP guide covers the paths to hosting it.
• A paid ChatGPT plan. Custom connectors require Plus, Pro, Business, Enterprise or Edu; on Business/Enterprise workspaces an admin may need to permit them. Free accounts get built-in apps only.
• Developer mode enabled — the beta toggle that unlocks full MCP tool support, including write actions.

That tester is at /tools/mcp-server-tester — thirty seconds there saves a debugging session inside ChatGPT’s less-verbose UI.

Once connected, ChatGPT lists the server’s tools and calls them as the conversation warrants, asking for confirmation before write actions. If tools don’t appear: confirm the URL points at the MCP endpoint itself (not your site root), that the server answers POST requests, and that any OAuth flow actually completed — community reports note connectors occasionally finishing OAuth without appearing in chat, in which case removing and re-creating the connector is the current fix.

Developer mode ships with unusually blunt warnings, and they’re worth taking at face value: custom connectors are not verified by OpenAI, and a malicious or compromised server can attempt to exfiltrate conversation data or abuse the write access you grant it.

• Connect only servers you control or trust. The threat isn’t hypothetical — poisoned tool descriptions are a documented attack class; our MCP prompt injection guide walks the mechanics.
• Scope what the server exposes. A server holding business data should enforce per-key permissions server-side — read-only keys, redacted fields — so a connected assistant can only reach what that key allows.
• Watch write actions. Review ChatGPT’s confirmation prompts rather than reflex-approving; an assistant that can send, spend or delete deserves the same scrutiny as a new hire with the same access.

Two more OpenAI surfaces speak MCP, and they matter if you’re building rather than just connecting:

The API platform. OpenAI’s developer platform supports remote MCP servers as a tool — an agent you build on OpenAI models can be pointed at the same MCP endpoint ChatGPT uses, with the platform handling discovery and tool calls. Details live in OpenAI’s platform documentation.

Apps in ChatGPT. The Apps SDK builds interactive ChatGPT apps *on top of* MCP — an app is an MCP server extended with UI components. If you’re weighing building an app versus a plain server, that trade-off has its own guide: OpenAI Apps SDK vs MCP.

The strategic takeaway is the reason this site exists: because ChatGPT, Claude, Gemini and the IDE clients all converged on MCP, one well-built hosted server makes you reachable from all of them — which is what MCP hosting is about.