Connect an MCP Server to ChatGPT: Fix the Missing + Button
If you have an MCP server URL and there is no plus button on ChatGPT’s Plugins page, nothing is broken. The button that adds a custom MCP server is gated behind Developer mode, and until that setting is on, it is not rendered at all. There is no error, no greyed-out control, and no message telling you what is missing. The page simply looks like a page that cannot do what you are trying to do.
The reason this wastes so much time is that Developer mode is not on the Plugins page, and it is not where most walkthroughs published before this summer say it is.
Where Developer mode actually lives now
Per OpenAI’s own instructions, the path is:
- Open Settings.
- Select Security and login.
- Turn on Developer mode.
That is worth stating plainly because a lot of writing on the web, including our own earlier guide, sends you to Settings → Connectors → Advanced. That path was correct when ChatGPT called this surface Connectors. It was renamed to Plugins in July 2026, and the Developer mode toggle went to Security and login. If you have been clicking through Connectors looking for an Advanced section, that is why you cannot find it.
The full setup
Once Developer mode is on, go back to the Plugins page and the plus button is there.
- Select the plus button.
- Enter a user-facing name and description. Use a name you will recognize in the tools menu later, because that is where you will go looking for it.
- Under Connection, paste the MCP server URL from your provider, including the
/mcppath. - Create the connection.
- Review the tools and metadata ChatGPT discovered from the server.
Step 3 is the one that fails silently. A server’s marketing page and its MCP endpoint are usually not the same URL, and a connection to the bare domain will not initialize. If your provider gives you something ending in /mcp, that trailing path is load-bearing.
Read the permission warning before you accept it
ChatGPT shows a security warning at creation time, and it is not boilerplate. An MCP server can expose tools that create, update, delete or send things on your behalf, not only tools that read. Some servers are deliberately read-only; others are not, and the connection dialog is the last point where you get to decide.
Three things worth doing before you click through it: confirm the URL came from the provider rather than a search result, check what the discovered tool list actually contains, and understand that approving the connection approves the whole tool surface, not one action.
Saved in settings is not the same as available in chat
The step people skip is the one that tells you whether any of it worked. Open a conversation, use the tools menu, and search for the plugin by name. A connection can sit in your settings looking healthy while the model never picks it up in a conversation.
If it is there, ask for something the server should obviously handle and watch which tool gets selected. Indirect phrasing is the better test: a request that expresses the goal without naming the tool tells you whether the descriptions on the server are good enough for the model to route to it on its own.
When the toggle is missing, not the button
These are different problems with different fixes.
| What you see | What it means |
|---|---|
| No plus button on Plugins | Developer mode is off. Turn it on under Security and login. |
| No Developer mode toggle in Settings | Account or workspace policy. On a managed workspace an admin controls this. |
| Connection created, no tools listed | The server responded but exposed nothing usable. Check it with MCP Inspector. |
| Connection fails outright | Endpoint, transport or auth. See below. |
OpenAI states that Developer mode availability “can depend on account and workspace policy.” If you are on a company workspace and the toggle is absent rather than off, that is an admin decision and no amount of clicking will surface it.
When the connection itself fails
The endpoint has to be reachable over public HTTPS and support streamable HTTP, typically at /mcp. A server running only on your laptop will not work, because ChatGPT calls it from OpenAI’s side rather than from your machine. For a private server, OpenAI’s Secure MCP Tunnel is the supported route.
To find out what is actually wrong before you keep retrying the dialog, list and call the server’s tools directly:
npx @modelcontextprotocol/inspector@latestThat separates the two possibilities cleanly. If Inspector cannot reach the server either, the problem is the server. If Inspector works and ChatGPT does not, the problem is transport, initialization, schema or authentication between them.
What you get once it connects
A connected MCP server is what turns ChatGPT from something that describes a task into something that performs it. Carly’s MCP is the example used in the walkthrough: connect your email accounts, calendars and file storage once, add the URL, and ChatGPT can act across all of them in a single request rather than one connected inbox. It reaches 260+ native integrations plus your own API key for anything else with an API.
One limitation is worth knowing before you build on it.
The unattended-work gap closed when ChatGPT and Claude Cowork added scheduled runs. Their standard schedulers still start from the clock, not arbitrary events in connected apps; API-triggered ChatGPT Workspace Agents require an external event detector and caller.
That is the gap Carly’s own workflows fill: they start on an event, an inbound email, a booking, a changed record, and run without anyone opening a chat window. Zapier-style workflow steps are free, and AI agents start at $35 a month. Connect your accounts at carlyassistant.com.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
