ChatGPT Apps and Connectors Not Showing or Not Working: Fixes

When a ChatGPT connector seems to be missing or broken, the most common cause is that it has moved: connectors are now called apps, and apps are installed as part of plugins from the Plugins tab, not from a connectors page in Settings. After that, work down a short list. Is the feature on your plan? Has your workspace admin allowed this plugin and its app for your role? For your own MCP server, is developer mode on? Has the app’s sign-in expired, or does your account lack access in the other service? Is the app turned on in this chat? Is it available where you are and on this device? And does the model you picked support the tool? Each check takes a minute, and the fix is usually a setting, not a bug.

What apps and plugins are, what they can reach and how to add one is covered in ChatGPT connectors and apps, and not repeated here. This post is the troubleshooting guide, following OpenAI’s documentation as it reads at the end of September 2026.

First: the rename

Older guides send you to Settings, then Apps & Connectors. That is out of date. OpenAI’s plugins documentation↗ (opens in a new tab) now describes one Plugins Directory, shared by ChatGPT and Codex, opened from the Plugins tab. A plugin can contain skills and MCP servers, and OpenAI uses app and MCP server for the same connected integration. So a “Google Drive connector” is today the Google Drive plugin, and the connection inside it is an app. If a guide’s screenshots do not match your screen, this is why.

Quick triage

1. Can you see the Plugins tab and the plugin you want? If not, check your plan, your platform and your admin.
2. Is the plugin installed, and does its page say Disabled by admin? If so, only an admin can change it.
3. Did ChatGPT ask you to connect an account, and did you finish? If the connection fails, reconnect.
4. Are you in a new chat started after installing it, with the plugin chosen or mentioned with @?
5. Does it fail only for writes? Look at the app’s action settings and the approval prompt.
6. Does it fail only with one model? Switch models and try the same request.

Cause 1: your plan or platform

• Symptom: the plugin, or the option to add your own server, is not there at all.
• Cause: availability depends on plan and surface. Plugins work in Chat and Work on the web, desktop and mobile, but the Codex IDE extension does not support them, and plugins marked Desktop only can be discovered on the web but only installed and used in the ChatGPT desktop app. Developer mode, which you need for your own MCP server, is listed for Pro, Plus, Business, Enterprise and Education accounts on the web.
• Fix: open the plugin’s details page to see where it works, try it on the surface it names, and check your plan before assuming a fault.

Cause 2: your workspace admin’s policy

• Symptom: the plugin shows Disabled by admin, installs but its app will not connect, or reads work while writes are refused.
• Cause: in Business, Enterprise and Edu workspaces access passes through several separate layers: whether the plugin is available to your role, whether you may use its MCP server, which actions are allowed, and what your account can do in the service itself. A request has to pass every one.
• Fix: tell your admin which layer failed, not just “it does not work”. The plugin, the app, the action, or your permissions in the service are different settings.

Defaults explain a lot of this. The ChatGPT Work overview for admins↗ (opens in a new tab) says plugins and their underlying apps are off by default for Enterprise and Edu workspaces and on by default for Business. It adds that making a plugin available does not enable its required app or grant access to an account, and that admins can restrict an app to read-only actions or an approved set. A write that fails while reads succeed is often exactly that restriction working.

Cause 3: developer mode for your own MCP server

• Symptom: you want to add a server that has no listing, and the plus button in Plugins offers nothing useful, or the app you created never appears in a chat.
• Cause: custom MCP servers need developer mode. The plus button only creates developer-mode apps after you turn it on, the apps land under Drafts, and they are used from the Developer mode tool in a chat, not from the ordinary plugin list.
• Fix: turn on Settings, Security and login, Developer mode. Create the app from Plugins with the server URL. In the chat, choose Developer mode from the plus menu and select the app for that conversation.

OpenAI’s developer mode guide↗ (opens in a new tab) supports SSE and streaming HTTP, with OAuth, no authentication or a mix, and notes that tools without a readOnlyHint annotation are treated as writes that ask for confirmation. In a workspace, developer mode availability can also depend on workspace policy, so a missing toggle on a work account is a question for the admin.

If the app is created but fails to connect, test the server outside ChatGPT. OpenAI’s guide to connecting and testing a plugin↗ (opens in a new tab) says the server must be reachable over public HTTPS, or through its Secure MCP Tunnel, and suggests the MCP Inspector for listing and calling tools directly:

npx @modelcontextprotocol/inspector@latest

If you changed the server’s tools and ChatGPT still shows the old ones, open the app in Plugins and select Refresh, then start a new conversation. Developer-mode apps do not pick up new tools, descriptions or server instructions until you do.

Cause 4: the sign-in has expired or lacks access

• Symptom: the app worked last week and now errors, asks you to connect again, or returns nothing for files you know exist.
• Cause: the connection’s token has expired or been revoked, or the account you signed in with cannot see that file, channel or record. An app reaches only what that account can reach; nothing in ChatGPT widens it.
• Fix: disconnect and reconnect the app, signing in with the right account. If results are still empty, check the permission in the other service itself.

Two details trip people up. Sign in with ChatGPT, where a partner offers it, shares only your name, email address and profile picture; you still approve the plugin’s permissions as a separate step. And uninstalling a plugin removes the bundle but leaves a separately connected MCP integration connected, so a “fresh install” may be reusing the old, broken connection until you disconnect it too.

Cause 5: the app is not on in this chat

• Symptom: the plugin is installed and connected, but ChatGPT answers from its own knowledge or from the web instead.
• Cause: bundled skills become available in new chats after installation, not in the chat you were already in. ChatGPT picks tools automatically when a request matches, and an installed plugin will not run on every request. Developer-mode apps are selected per conversation.
• Fix: start a new chat, type @ and choose the plugin, then name the app and the tool you want. OpenAI’s own advice for developer mode is to be explicit and to say which other tools not to use.

Approvals reset too. You can remember an approve or deny choice for a tool, but only for the rest of that conversation; a new conversation asks again.

Cause 6: region and device

• Symptom: a colleague in another country, or on another device, sees an app or feature that you do not.
• Cause: OpenAI rolls some features out by region and by platform, and its documentation marks some features, such as Computer Use in the desktop app, as available only in supported regions.
• Fix: check the plugin’s page and OpenAI’s help centre for where it is offered, and try the platform the page names before reporting a fault.

Cause 7: the model in use

• Symptom: the same request calls the tool with one model and ignores it with another, or a saved task starts failing.
• Cause: tool support can differ by model. OpenAI’s What’s new digest↗ (opens in a new tab) gives an example: Site tools are not available with GPT-5.6 Luna. The same digest says GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on 14 October 2026, so saved settings, scheduled tasks and custom agents that still select it need changing.
• Fix: switch to another model and repeat the request. If it works, keep that model for tool-heavy chats.

When the server is a task board

fenbs is a small task board that ChatGPT reaches as a developer-mode app, where your plan allows it, at https://fenbs.ai/api/mcp. The same checks apply. The app must be created with developer mode on and selected in the chat. The sign-in is OAuth in your browser, and ChatGPT then works under your role on the board, narrowed by the scopes you approve. If you revoke that access under Settings in fenbs, ChatGPT’s next call fails until you reconnect, which is the intended behaviour. Every task it adds, moves or comments on is recorded in History under its own name, so when a write goes wrong you can see which assistant made it.