# ChatGPT MCP: How to Add a Server (2026)

> Add an MCP server to ChatGPT with Developer Mode — enable it, create a custom connector with your server's HTTPS URL, and call its tools in chat.

Source: https://designrevision.com/blog/add-mcp-server-to-chatgpt

---

To add an MCP server to ChatGPT, enable **Developer Mode** and add the server as a **custom connector** with its HTTPS URL. Since late 2025, ChatGPT has been a full [Model Context Protocol](https://modelcontextprotocol.io) client — it can call an MCP server's tools for both read *and* write actions — so the same servers you use in Claude Code or Cursor work here too. The one rule that trips people up: ChatGPT connects to **remote** servers over HTTPS, not local stdio ones. This guide covers the exact steps, the plan requirements, and how to wire up a real server.

*Last updated: July 2026. ChatGPT's MCP support is a fast-moving beta; the Developer Mode flow below was verified against OpenAI's docs in July 2026 — check [OpenAI's help center](https://help.openai.com/en/articles/12584461-developer-mode-and-mcp-apps-in-chatgpt) if a menu label has moved.*

## First, Which "ChatGPT MCP" Do You Mean?

Two things get called "ChatGPT MCP," and it's worth a one-line disambiguation:

- **Adding MCP servers *to* ChatGPT** — making ChatGPT a client that calls external tools. That's this guide.
- **Building an MCP server *for* ChatGPT** — packaging your own app as a **ChatGPT MCP server** (a "ChatGPT App") so others can use it via the [OpenAI Apps/MCP docs](https://developers.openai.com/api/docs/mcp). Different job, different audience.

If you want ChatGPT to reach your repos, database, or component library, you're in the right place.

## Prerequisites

- **A paid plan.** Developer Mode and MCP connectors are available on ChatGPT **Plus, Pro, Business, Enterprise, and Edu** — not the free tier. On Business/Enterprise/Edu, an admin may need to enable connectors for the workspace first.
- **A remote MCP server (HTTPS).** ChatGPT speaks **HTTP/SSE**, so the server must be reachable at a public HTTPS URL. A local stdio server won't work as-is — you'd tunnel it first (more on that below).

## Step 1: Enable Developer Mode

Developer Mode is the switch that turns on full MCP client support (all tools, read and write). To enable it:

1. Click your **profile icon** (top-left in ChatGPT) and choose **Settings**.
2. Open **Apps & Connectors**.
3. Scroll to the bottom and open **Advanced Settings**.
4. Toggle **Developer Mode** on.

That's the beta that gives ChatGPT "full Model Context Protocol client support for all tools, both read and write" — with an explicit confirmation before any write action.

## Step 2: Add a Custom MCP Connector

With Developer Mode on, go to **Settings → Apps & Connectors → Create** and fill in:

| Field | What to enter |
|-------|---------------|
| **Name** | A friendly label, e.g. `GitHub` |
| **Description** | One line on what it's for |
| **Connector URL** | The server's public **HTTPS** MCP endpoint (e.g. `https://…/mcp`) |
| **Authentication** | `None`, `OAuth`, or `Token` — whatever the server requires |

Click **Create**. If the server uses OAuth, ChatGPT runs the sign-in flow; if it uses a bearer token, choose **Token** and paste it. The connector's tools then register automatically.

## Step 3: Use the Tools in a Chat

Once a connector is live, you don't pick tools manually — **ChatGPT decides which to call from your request.** Ask it to do something the server enables ("open a PR for this fix," "pull the latest rows from the users table") and it invokes the matching tool, confirming before any write. You can wire several connectors at once and ChatGPT will orchestrate across them in a single task.

## Add DesignRevision MCP to ChatGPT

Because [DesignRevision MCP](/mcp) — the server that installs real [shadcn/ui components](/components) — is already a **remote** server, it's a direct fit for ChatGPT's remote-only model: no tunnel, no bridge. In **Settings → Apps & Connectors → Create**, enter:

- **Name:** `DesignRevision`
- **Connector URL:** `https://mcp.designrevision.com/mcp`
- **Authentication:** `Token` → paste your DesignRevision token

Create it, and ChatGPT can discover, inspect, and install shadcn/ui components on request. It's freemium, so you can connect it and pull components on the free plan.

## Developer Mode vs. Deep Research Connectors

ChatGPT has had a narrower "connectors" concept for a while, and it's easy to conflate the two:

- **Deep Research connectors** only require a server to expose two **read-only** tools — `search` and `fetch` — and are limited to retrieval for research answers.
- **Developer Mode** is the full MCP client: **every** tool a server exposes is available, including write actions. This is what you want for anything beyond search.

If a server's tools aren't showing up, confirm you're in Developer Mode, not just using it as a Deep Research source.

## What About Local Servers?

ChatGPT can't launch a local stdio server the way [Claude Code](/blog/add-mcp-server-to-claude-code) or [Cursor](/blog/add-mcp-server-to-cursor) do. If the server you want only runs locally, expose it on a public HTTPS URL first — a tunnel such as ngrok in front of the local process, then add that URL as the connector. Servers that are already hosted remotely (like DesignRevision MCP) skip this entirely. For how every agent handles this differently, see our [comparison of MCP clients](/blog/mcp-clients-compared).

## Troubleshooting: When a Connector Won't Work

- **Connector created but no tools show up.** Confirm **Developer Mode** is actually on (Settings → Apps & Connectors → Advanced Settings). Without it, ChatGPT treats the server as a Deep Research source and only surfaces `search`/`fetch`. Reopen the chat after adding a connector so the tool list refreshes.
- **OAuth vs. Token — which do I pick?** Match the server. If it hands you a sign-in flow (most hosted vendor servers), choose **OAuth**; if it authenticates with a bearer token or API key, choose **Token** and paste it. Guessing wrong is the usual cause of a connector that saves but then returns 401s.
- **"Failed to connect" or 401 errors.** The URL isn't a reachable HTTPS MCP endpoint, or the token is wrong or expired. Open the `/mcp` URL in a browser to confirm it responds, and regenerate the token if in doubt — remember ChatGPT needs a public HTTPS server, so a `localhost` URL will never connect.
- **Tools appear but never fire.** ChatGPT decides when to call a tool from your phrasing, so be explicit about the action ("use the GitHub connector to open a PR"). Writes also pause for a confirmation you have to approve before anything happens.

## A Note on Safety

Developer Mode can take real actions, so treat connectors like installed software. OpenAI surfaces a confirmation before writes, but the safer habit is to add servers you trust — official vendor servers or ones you host — review what each can do, and keep secrets out of connectors you didn't set up. The convenience of write-capable MCP is exactly why it deserves a moment's caution.

## Which MCP Servers Should You Add?

The servers are the same across agents; only the way you add them changes. A focused set of GitHub, Context7, a database server, and **DesignRevision MCP** for UI covers most work. We rank the best picks — each of which works in ChatGPT as a remote connector — in our roundups for [Claude Code](/blog/best-mcp-servers-for-claude-code), [Cursor](/blog/best-mcp-servers-for-cursor), and [Codex](/blog/best-mcp-servers-for-codex).

## Conclusion

ChatGPT is a genuine MCP client: flip on **Developer Mode**, add a **custom connector** with the server's HTTPS URL and auth, and its tools show up in chat — read and write, with a confirm step. The only real constraint is remote-over-HTTPS, which is why an already-remote server like DesignRevision MCP connects in one step while a local server needs a tunnel first. Add the two or three servers that fill a real gap, and keep the set to ones you trust.

---

## Related Resources

- [DesignRevision MCP — the shadcn/ui MCP server](/mcp)
- [MCP Clients Compared: Every AI Coding Agent](/blog/mcp-clients-compared)
- [Perplexity MCP: The Complete Guide (2026)](/blog/perplexity-mcp)
- [How to Add an MCP Server to Claude Code (2026 Guide)](/blog/add-mcp-server-to-claude-code)
- [How to Add an MCP Server to Cursor (2026 Guide)](/blog/add-mcp-server-to-cursor)
- [How to Add an MCP Server to Codex CLI (2026 Guide)](/blog/add-mcp-server-to-codex)
- [Browse the shadcn/ui component registry](/components)
