Apr 26, 2026

Last updated on Oct 1, 2026

Using MCP Servers in ChatGPT: Developer Mode + Apps SDK

ChatGPT supports MCP now — the same open protocol Anthropic introduced in late 2024 — which means the same servers that power Claude can power ChatGPT. This guide walks through enabling Developer Mode, connecting a custom MCP server, and using Apps SDK integrations, with Carly as the working example. If you’ve been waiting for ChatGPT to catch up on tool use, this is it.

Looking beyond ChatGPT? MCP servers by category covers the rest.

What ChatGPT’s MCP support actually does

MCP is a standard for letting an AI client call external tools — calendar, CRM, GitHub, your internal database, anything. ChatGPT can now connect to MCP servers and call those tools mid-conversation, which is what Claude has been doing in Claude Desktop and Claude Code for over a year. The one meaningful difference: ChatGPT connects to MCP servers over the network, and never launches one itself. It won’t start a local stdio server (the node my-server.js kind that Claude Desktop and Cursor run on your laptop). So you either need a hosted HTTPS endpoint, or you run the local server behind OpenAI’s Secure MCP Tunnel, a small client that sits on your machine or inside your network and opens an outbound connection ChatGPT can reach, with no public URL.

Plan availability

Custom MCP connectors live behind ChatGPT’s Developer Mode, which is available on Plus, Pro, Business, Enterprise, and Edu plans. Free ChatGPT does not support custom connectors. Apps SDK is the separate, developer-facing path: it’s how teams build integrations that show up in ChatGPT’s official app catalog for any user to install — no Developer Mode required on the user side. If you’re a builder, OpenAI’s MCP docs cover both paths and the Apps SDK examples repo has working code.

Two ways to connect

There are two flavors of “MCP in ChatGPT” and they get conflated all the time:

Custom Connectors / Developer Mode — for individuals adding an MCP server to their own ChatGPT account. You paste in a URL, OAuth or paste a token, and the tools show up in your chats. Good for personal workflows and trying servers nobody has packaged yet.
Apps SDK — for developers publishing an MCP-backed integration that any ChatGPT user can install from the app catalog. Apps SDK uses MCP under the hood but adds UI components, OAuth scaffolding, and review/distribution.

We’ll walk through Custom Connectors first because that’s what most people are searching for.

Step-by-step: enable Developer Mode in ChatGPT

Open ChatGPT (web or desktop).
Click your profile picture → Settings.
Go to Security and login.
Toggle Developer mode on.

You’ll get a warning that custom connectors run third-party code on your behalf. Read it, accept it, and you’re in. If there’s no toggle under Security and login but Create MCP App already shows in the Plugins Add menu, your account has Developer mode on and you can skip this section.

Step-by-step: add the MCP server

Open Plugins in the sidebar and click Add next to the search box.
Choose Create MCP App. The other two options, Create plugin and Upload plugin archive, are for packaged plugins; ChatGPT Plugins covers those. If Create MCP App isn’t on the menu, Developer mode is still off.
Give it a name, plus an optional icon and description.
Under Connection, paste the server’s HTTPS URL. For a private server, switch to Tunnel and pick the tunnel you created in OpenAI’s platform settings.
Set Authentication to OAuth, No Auth, or Mixed. OAuth is the default and what most hosted servers use.
Tick I understand and want to continue and click Create.

The Developer mode toggle moved. Until the July 2026 rename, Developer mode lived under Connectors → Advanced, and Connectors is now called Plugins. If you followed the old path and found no Advanced section, that is why. ChatGPT Developer Mode covers where it lives on each plan. Connecting an MCP server to ChatGPT walks the current flow end to end, including what to do when the toggle itself is missing rather than the button.

What MCP still won’t do for you

Once you have connectors working, the ceiling becomes obvious fast: MCP is request-response. You ask, a tool answers. Nothing in an MCP server can start a conversation.

That means the jobs people actually want automated stay manual. “When a customer replies, draft the follow-up and send it.” “When the deploy goes red, tell the on-call engineer.” “Every Friday, pull the numbers and email the summary.” None of those are expressible in MCP, because MCP has no way to notice anything happened.

MCP is only starting to learn how to start things. ChatGPT can now kick off an automation from an MCP event, but it is building on a proposal: the protocol's own Triggers and Events working group still lists the specification as ideating, and every app has to ship its own event support. OpenAI's docs still list only three apps whose events can start a task: Gmail, Slack and GitHub. Carly workflows start on their own when an email comes in, a meeting gets booked, a lead changes stage in the CRM or a row lands in a spreadsheet, and Carly can keep an eye on any app with an API.

That gap is what Carly is built for, and it’s worth being precise about how the two relate.

Carly is a plugin in ChatGPT’s own directory, so it skips the Developer mode setup above entirely. Click Get the Carly plugin, add it, and sign in. The Carly plugin brings all of your inboxes, calendars, files and CRM into the same chat, and Carly connects to thousands of apps. Setup for Claude Code, Cursor, VS Code and other clients is on Carly’s MCP page.

What Carly adds is the layer MCP doesn’t cover:

Workflows that fire on events — an email arrives, a form is submitted, a record changes, a schedule hits, and the work runs. No one has to open a chat.
It sends, rather than drafts — real email through Gmail and Outlook, from your address, with attachments.
Multi-step across apps — one trigger can update a CRM, book a calendar slot, post to Slack, and reply to the customer as a single flow.
Agents with their own email addresses — you can CC one on a thread and it participates, rather than you copying output out of a chat window.
Free plumbing — Zapier-style workflow steps that don’t use AI run free, and AI agents start at $35/month.

So the honest framing: use MCP connectors to give ChatGPT reach inside a conversation. Use Carly for the work that has to happen when you aren’t in one.

Best MCPs to try in ChatGPT right now

Most of the MCP servers worth using in ChatGPT today are ones that already publish hosted/remote endpoints. A starting set:

Cloudflare — official remote MCP servers for Workers, R2, KV, and observability. Hosted by Cloudflare.
Stripe — official Stripe MCP server for payments, customers, invoices.
Linear — official Linear MCP server, OAuth-based, great for “what’s blocking my team this week.”
GitHub — GitHub’s hosted MCP server covers issues, PRs, code search.
Notion, Asana, Intercom, Sentry — all ship official remote MCP servers in 2025–2026; pull the URLs from each vendor’s docs.

For a wider catalog, see the official MCP server registry. And for the protocol itself, what is MCP? covers the architecture without the install commands.

Limitations and gotchas

A few things will trip you up:

Local stdio servers need a tunnel. If a server’s setup says command: "node" or command: "npx", ChatGPT can’t launch it. Run it behind a Secure MCP Tunnel or wait for a hosted version.
OAuth happens inside ChatGPT. Servers that use OAuth will prompt you in a popup the first time. Make sure popups aren’t blocked, and re-auth if your token expires.
Tools aren’t available in shared chats. Custom connectors are scoped to your account. If you share a conversation, the recipient won’t be able to re-run the tool calls.
Rate limits apply. ChatGPT enforces its own per-tool rate limits on top of whatever the MCP server enforces. High-volume agentic loops will hit ceilings faster than they would in Claude Code.
Team/Enterprise admin controls. On Team and Enterprise plans, admins can disable Developer Mode org-wide or allowlist specific connectors. If your toggle is missing, it’s likely a workspace policy.
Apps SDK ≠ Custom Connectors. If you’re trying to publish a connector for other people to use, you want Apps SDK and its review process — pasting your URL into a friend’s ChatGPT works, but it’s not how you ship to the catalog.

For installing MCP in Claude, Cursor, or Windsurf instead, see the install guide.

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."

Gus Ibrahim, Founder & Director, IHR