Back to Guides
/ Guides

How to Connect ChatGPT to an MCP Server (Developer Mode, 2026)

Aleks Koha6 min read

Yes, ChatGPT can use MCP servers — full ones, with real tools, including write actions. The reason so many people believe otherwise is that the capability hides behind a toggle called developer mode, and the Reddit threads asking "is MCP ever coming to ChatGPT" outrank the documentation saying it already did.

This guide is the short version: where the toggle is, how to add a custom connector, and a worked example using a remote server that turns ChatGPT from a chat window into something that can provision and run real things.

Can ChatGPT Use MCP Servers?

ChatGPT supports custom MCP connectors through developer mode. Once enabled, you can register any remote MCP server, and every tool that server exposes — read and write — becomes available in your conversations, subject to confirmation settings. Supported transports are Streamable HTTP and SSE; authentication can be OAuth or none.

The distinction that trips people up: ChatGPT's regular "connectors" (Google Drive, and friends) are curated, search-and-fetch integrations. Developer mode is the door for arbitrary MCP servers — yours, or any remote endpoint you trust.

What You Need First

Developer mode is available on Pro, Plus, Business, Enterprise, and Education accounts, on the web app. You'll also need the URL of a remote MCP server. For the worked example we'll use the Matagi MCP server:

https://mcp.matagi.ai/mcp

It's a single remote endpoint over Streamable HTTP with OAuth — no API keys to generate, nothing to install or self-host, which makes it a clean first connector even if you just want to see MCP working.

Step 1: Enable Developer Mode

In ChatGPT on the web, go to Settings → Apps → Advanced settings and switch on Developer mode.

On Business and Enterprise workspaces, an admin may need to allow it first: Workspace Settings → Permissions & Roles → Connected Data → Create custom MCP connectors. If you don't see the toggle at all, that's usually why.

Step 2: Create the Connector

Go to Settings → Connectors and click Create. You'll be asked for:

Name. User-facing — Matagi works.

Description. This matters more than it looks: the model reads it when deciding whether to use the connector. Something like "Provisions infrastructure and runs agents — use for building, deploying, databases, and scheduled jobs."

MCP server URL. https://mcp.matagi.ai/mcp

Authentication. Select OAuth. ChatGPT discovers the server's OAuth metadata and registers itself automatically — that's dynamic client registration doing its job, and it's why there's no "paste your API key" step.

Save the connector.

Step 3: Authenticate and Use It

The first time you use the connector (or when you connect it explicitly), a browser consent screen opens. Sign in with your work email and approve access. If you're new to Matagi, your workspace is created on the spot with trial credit.

Back in the chat, add the connector to a conversation (via the plus/tools menu) and just describe what you want. Tool calls show up inline; write actions ask for confirmation before executing, which is exactly the behavior you want while you're building trust.

If you later change or update a connector's tools, open it under Settings → Connectors and hit Refresh to re-pull the metadata.

What to Actually Connect

The MCP ecosystem is full of servers that make ChatGPT better at reading things. Useful, but the ceiling is low: a docs server makes answers more accurate; it doesn't change what ChatGPT can do.

The Matagi server is on the other end of the spectrum. Connected, ChatGPT can provision serverless functions, isolated Postgres databases, batch jobs over CSVs, and deployments; register cron schedules that keep running after your chat ends; and stand up always-on agents wired to Slack, Telegram, email, or web chat. Anything it builds can call every major model and 3,000+ external tools through Matagi's proxy, with credentials injected server-side — never pasted into a chat, never sitting in generated code.

Concretely: "build me a form endpoint that stores submissions in a database and emails me a daily summary" is, with this connector, a thing ChatGPT can finish — not just draft the code for. And because the workspace lives behind the MCP endpoint rather than in the chat, the same resources are reachable later from Claude, Cursor, or Codex. Nothing is stranded in one vendor's client.

Security Notes Worth Reading

Developer mode is powerful precisely because it's unrestricted, so apply judgment:

Only connect servers you trust. A malicious MCP server can attempt prompt injection or data exfiltration. Prefer endpoints with real OAuth, scoped workspaces, and audit trails over anonymous community servers for anything touching your data.

Keep confirmations on for write actions until a connector has earned autonomy. Model mistakes on writes are rare but expensive.

Don't paste secrets into chat. With a proxy-based setup like Matagi's, you shouldn't need to — credentials for external tools live server-side, encrypted and revocable, with a full audit log.

The recurring rule from everything else we've written applies here too: automate the volume, keep the judgment. Let the connector do the provisioning; you review what it built before it matters.

FAQs

Is MCP available in the ChatGPT desktop and mobile apps? Developer mode setup lives in the web app. Availability of custom connectors on other surfaces has been expanding — configure on web first, then check your other clients.

Which plans support custom MCP connectors? Pro, Plus, Business, Enterprise, and Education. On managed workspaces an admin has to permit custom connectors.

Does ChatGPT support local MCP servers? Custom connectors expect a remote endpoint (Streamable HTTP or SSE). To use tools that only exist as local servers, host them — or use a platform whose server is already remote, like Matagi.

What's the difference between this and ChatGPT plugins/apps? Apps built on the Apps SDK are packaged, reviewed experiences. A custom MCP connector is raw tool access for your own workspace — no review process, no packaging, available immediately.

Can I use the same MCP server in Claude and Codex too? Yes — that's the appeal of a remote endpoint. Matagi's URL works from Codex, Claude, Cursor, and any MCP client, all landing in the same workspace.

Try it with a server that can actually build: connect the Matagi MCP endpoint to ChatGPT's developer mode and describe what you want shipped. New workspaces include trial credit.

chatgpt mcpmcp connectordeveloper modemcp serverschatgpt

Build your first AI agent free

Describe what you want done in plain English. Matagi provisions the infrastructure, wires the integrations, and deploys it.

Get started