ChatGPT Developer Mode: What It Is and How to Enable It

ChatGPT developer mode is a beta setting that lets ChatGPT use a full MCP client: it can call any tool a remote MCP server offers, including tools that change data, not only read and search. You turn it on in settings, add a server URL, and approve actions as they come up. Availability depends on your plan and workspace admin.

What is ChatGPT developer mode?

Normal ChatGPT connectors are built for reading. They fetch and search information from an app and hand it back to the chat. Developer mode removes that limit. OpenAI describes it as support for apps with full Model Context Protocol (MCP) capability, including write and modify actions. That means ChatGPT can create a record, update a row, send a message or run any other tool the server exposes.

In practice, developer mode does three things:

• It lets you add your own remote MCP server by URL.
• It lets ChatGPT call every tool that server lists, not just a search and fetch pair.
• It lets you choose which apps are active in each conversation.

The feature is labelled beta, so functionality, interface and permissions may change. Treat any menu names below as a snapshot of October 2026 and check them against your own screen.

Who gets developer mode?

This is the part sources disagree on, so check your own account. OpenAI's help article lists Business, Enterprise and Edu workspaces on the web; some guides also report Plus and Pro. In a workspace an admin may need to switch developer mode on before members can use it.

Older articles describe the same switch under a different menu name. OpenAI's own wording is the reference, and the quickest test is simple: open your settings and look for the Developer mode toggle. If it is missing, ask a workspace admin whether it is enabled, or check the plan documentation on OpenAI's help site. It also appears on ChatGPT web only, so do not expect it in the mobile app.

How do you enable developer mode in ChatGPT?

These are the steps OpenAI describes; menu names may vary. They follow OpenAI's Connect from ChatGPT page:

1. Turn on developer mode if needed. Open Settings, then Security and login, and switch on Developer mode. In a Business, Enterprise or Edu workspace an admin may have to allow it first.
2. Open Plugins. In ChatGPT, open Plugins and select the plus button.
3. Choose Add custom MCP server.
4. Enter the details. Give it a name and a description, then paste the public server URL, including the /mcp path, as the Connection.
5. Set authentication. Choose none if the key is already in the URL; otherwise pick what the server needs, such as OAuth.
6. Accept the risk warning and select Create.
7. Use it in a chat. In a new conversation, type @ in the prompt box and select the app. Older flows use the plus menu, then Developer mode.

Older accounts may still show Settings, then Connectors, then Advanced, then Add connector instead.

If you only want to read from a server and do not need write tools, you may not need developer mode at all. The guide to connecting an MCP server to ChatGPT covers the standard connector route, and the MCPifex ChatGPT page shows the short version.

What changes compared with normal connectors?

The big change is that ChatGPT can act, not just look. A read-only connector can return a search result. A developer mode app can also run whatever write tools the server publishes. If a database server has an execute tool, ChatGPT can call it. If a messaging server can send, ChatGPT can send.

AreaNormal connectorDeveloper mode app

Tools available

Read and search

Every tool the server lists, including writes

Who adds a server

Picked from a list or approved by an admin

You, by pasting a URL

Review

Vetted before it is offered

Unreviewed: you are the reviewer

Confirmations

Rarely needed, nothing changes

Asked before actions, depending on permissions

How do confirmations work?

OpenAI says ChatGPT asks you to confirm based on the app's permissions and the context of the action. It does not promise a prompt for every call, and it says some especially risky actions may be blocked instead of offered for approval. Read each confirmation before you accept it. Check the tool name and the arguments, because a prompt that looks routine can still change real data.

Confirmations are a last line of defence. They do not replace limiting which tools exist in the first place.

How do you add a remote MCP server in developer mode?

You need a remote MCP server reachable over HTTPS. A server that only runs on your laptop over stdio will not work, because ChatGPT connects from OpenAI's side. The remote MCP server entry explains the difference. Your options are to host one yourself, use a vendor's remote server, or use a hosted service.

One example: with MCPifex you create a server instance in the portal, save your credentials, choose which tools are allowed and copy a key. ChatGPT's custom MCP form has no header field, so the key goes in the URL path instead. The connector URL then looks like this, with your own key in place of the placeholder:

https://mcp.mcpifex.com/mcp/YOUR_API_KEY_HERE

Paste that as the Connection when you add a custom MCP server, then create it. Treat the full URL like a password, since anyone who has it can use the instance. The MCPifex ChatGPT docs walk through the exact fields.

Whichever server you use, first confirm it works. Open a new chat, type @ in the prompt box, select the app and ask for something harmless and read-only, such as listing tables or searching a record. If ChatGPT lists the tools the server exposes, the connection works.

How do you use developer mode safely?

Developer mode hands a language model real write access. OpenAI cautions that custom servers are not reviewed by OpenAI, so connect only servers you trust: a malicious or compromised server can manipulate what the model does with your data. The practices below keep the risk small.

• Start with writes off. Expose only read tools first. Add a write tool when a specific task needs it, then switch it off again afterwards.
• Use a tool allow-list. Allow the few tools you need instead of everything. Fewer tools means fewer ways for a mistaken call to do damage. See how to control what AI agents can reach.
• Use least-privilege credentials. Give the server a database role or API key that can do only what the task needs. A read-only role makes a write tool harmless.
• Watch for prompt injection. Text returned by a tool, such as a web page or a message, can contain instructions aimed at the model. Read prompt injection through MCP tools and never auto-approve actions that follow a tool result.
• Read every confirmation. Check which tool will run and what it will change before you click through.
• Keep a log. Know what was called and when, so you can undo a mistake. The MCP security checklist covers logging and key rotation.
• Use a test account first. Point a new write-capable app at staging data, not production.

MCPifex fits this pattern because write and destructive tools start off. For PostgreSQL, query is available and execute stays off until you turn it on. For App Store Connect, the 38 write tools are off by default.

Why isn't developer mode working?

Most failures come down to a handful of causes.

• No Developer mode toggle. Your plan may not include it, or a workspace admin has not enabled it. Ask the admin, and check that you are on ChatGPT web.
• No custom MCP server option. It may appear only after developer mode is on. Reload the page after switching it.
• The server will not connect. Confirm the URL is public HTTPS, ends at the correct path and uses Streamable HTTP. A local server will not work.
• Authentication fails. Check that the key or OAuth setup matches what the server expects, and that the key has not been rotated.
• ChatGPT sees no tools. The server may expose none, or a policy may have disabled them all. Check your allow-list.
• The app is not used in the chat. Developer mode apps are active per conversation. In a new conversation, type @ in the prompt box and select the app. Older flows use the plus menu, then Developer mode.
• A tool call is refused. It may be blocked by the server's policy or by ChatGPT as too risky. That is expected behaviour, not a bug.

If a server connects in another client but not in ChatGPT, compare the transport and the URL first. The list of AI clients that support MCP notes which client supports which setup.

Sources

1. OpenAI Help Center, "Developer mode and MCP apps in ChatGPT (beta)".
2. OpenAI Developers, "Connect from ChatGPT".
3. Model Context Protocol, "Architecture".
4. Model Context Protocol, "Tools" specification.

← Back to blog