How to Use MCP Servers With ChatGPT

ChatGPT can use MCP servers, with two catches that trip most people up. First, it only connects to hosted servers reached over a URL: you can't point it at a command on your laptop the way you can with Claude Desktop or Cursor. Second, connecting your own server needs developer mode, and what you can do with it depends on your plan.

This guide covers who can use it, how to add a server, and what to do when the server you want only runs locally.

Which plans can connect MCP servers

On Business, Enterprise and Edu, admins and owners control which apps the workspace can use, and only they can publish an app for everyone.

Add a hosted MCP server

1. Turn on developer mode. On Business, go to Settings → Apps → Advanced settings and switch on Developer mode. On Enterprise and Edu it's under Settings → Apps → Advanced settings, if your admin allows it. Admins can also create apps for the whole workspace from Workspace settings → Apps → Create.
2. Create the app. Give it a name and paste the server's endpoint URL, for example https://mcp.linear.app/mcp for Linear.
3. Choose authentication. Most vendor-hosted servers use OAuth, so pick OAuth and sign in with your account on the vendor's page. Servers that need no sign-in can skip this.
4. Use it in a chat. Enable the app for the conversation and ask for something that needs it.

Every server in our MCP server directory that has a hosted endpoint lists the exact URL on its page, along with how it authenticates. Those are the ones you can connect to ChatGPT directly.

When the server only runs locally

Many MCP servers are local programs started with npx, uvx or docker. ChatGPT can't launch those. You have three options:

• Check for a hosted version. More vendors now run their own hosted server, and it's usually the better choice anyway: no install, and the vendor keeps it updated. The directory page for each server says whether a hosted endpoint exists.
• Use OpenAI's Secure MCP Tunnel for servers that run on a private network, on-premises or on a developer's machine. It gives ChatGPT a route to the server without exposing it to the public internet.
• Use a client that runs local servers, such as Claude Desktop, Claude Code, Cursor or VS Code.

Be deliberate about write access

Developer mode is powerful because a connected server can act, not just read: update a record, send a message, change a setting. OpenAI warns that connecting untrusted servers raises the risk of prompt injection, where content the model reads tries to steer it into doing something you didn't ask for.

Three habits keep this manageable:

• Prefer servers maintained by the vendor over community ones. Every directory page says who maintains the server.
• Grant the narrowest access the vendor allows, and use read-only where that's enough.
• Review what ChatGPT is about to do before approving an action that writes, sends or deletes.

For a business, those habits belong in policy rather than memory. Our guide to what an SME AI policy should include covers the rules, and is ChatGPT safe for business data covers which tiers keep your data out of training.

ChatGPT or Claude for MCP?

Both now support hosted MCP servers well. The practical differences are local servers (Claude Desktop, Claude Code, Cursor and VS Code run them; ChatGPT doesn't) and plan limits (ChatGPT's write-capable MCP is on its business tiers, while Claude's custom connectors are on Pro and above). If your team already works in one, connect the servers there rather than switching tools. MCP is the same protocol either way, which is the point of it. Our plain-English guide to MCP explains why.