ChatGPT Connectors Not Showing? Connectors vs Plugins Explained

You open ChatGPT, head to settings, and the Connectors entry you used last spring is gone. Gmail still works in one chat, Google Drive is missing in another, and a teammate swears the same menu shows up on their laptop. In most cases nothing is broken. OpenAI renamed and moved the integrations area twice in under a year, and hard limits on plan, device and workspace settings decide who sees what.

This article sorts out the names, shows where each piece lives today, and gives you a fix order that takes about ten minutes. Along the way you will see how connectors, apps, plugins and custom GPTs relate, because the answer to "why can't I see it" usually hides inside that vocabulary.

💡 Quick answer: Since July 9, 2026, new integrations are added from the Plugin Directory at chatgpt.com/plugins. Existing app connections keep working, but the standalone Connectors menu entry is gone, and what you can use depends on your plan, region, device and workspace admin settings.

Menu labels below reflect what was documented in early October 2026. OpenAI changes these screens often, so treat the exact wording as a pointer, not a promise.

What Changed in ChatGPT's Menus

The confusion has a simple cause: one idea has carried several different names in three years. Each rename left a trail of outdated blog posts, forum threads and support answers, and many of them are still easy to find in search results. Follow the timeline and the missing menu stops being a mystery.

Plugins Came and Went

OpenAI launched the first ChatGPT plugins on March 23, 2023. They were API extensions that let the chatbot browse, calculate and book things through third parties. New plugin conversations stopped on March 19, 2024, and existing ones stopped working on April 9, 2024.

GPTs and GPT Actions took over as the official successor. A GPT bundles instructions and files, while an Action lets that GPT call an outside REST API. So when someone says "plugins" in a 2024 article, they usually mean this older, closed system.

Connectors Became Apps

In 2025 OpenAI adopted the Model Context Protocol (MCP), the open standard Anthropic released in November 2024, and shipped connectors that link ChatGPT to work data such as Gmail, Google Drive and Microsoft 365. Business workspaces got app support in a November 2025 preview, and in December 2025 connectors were renamed apps.

Same plumbing, new label. That rename is why older tutorials send you to a menu that no longer matches your screen.

The Plugin Directory Arrives

On July 9, 2026, OpenAI replaced the App Directory with the Plugin Directory, which serves both ChatGPT and Codex. According to OpenAI's help center page for apps, existing app connections are unaffected. What moved is the menu: new integrations now come from chatgpt.com/plugins, where a plugin is a package that can hold skills, apps and app templates.

So the word "plugin" is back, but it no longer means 2023 plugins. It now names the wrapper, and the connection inside it is still called an app.

Connectors vs Plugins at a Glance

Here is the vocabulary side by side. Keep this table open while you read forum threads, because half of the confusion comes from people using these words interchangeably.

What a Connector Actually Is

A connector is a permissioned link between ChatGPT and another service. You sign in through OAuth, approve what ChatGPT may read, and often approve what it may change. Under the hood it is an MCP server: a small program that exposes tools such as "search my drive" or "create a calendar event" in a format any compatible assistant can call.

That last point matters. Because MCP is a shared standard, the same server can work in ChatGPT and in other assistants. The connector is the connection. The assistant is just one client.

What a Plugin Means Now

Today a plugin is the packaged workflow. It can include skills (written procedures for a recurring task), one or more apps that reach outside services, and app templates that admins roll out across a team. Think of the app as the pipe and the plugin as the pipe plus the instructions for using it well.

💡 Rule of thumb: If the thing talks to an outside service, it is an app, which is the new word for a connector. If it also ships instructions for a whole workflow, it is a plugin.

Seven Reasons Connectors Go Missing

Most "not showing" reports trace back to one of seven causes. Check them in this order, from the quickest to rule out to the slowest.

Account and Plan Limits

• 1. Your plan. The Plugin Directory is visible on every plan, but what works depends on your subscription tier. Reports from September 2026 list plugins as included on Free, Go, Plus and Pro, with some plugins greyed out when your tier does not qualify.
• 2. Your region. Some apps are not offered in every country. A greyed-out button usually means plan or region, not a bug.
• 3. The wrong account. People with a personal login and a work workspace often look in one while the connector lives in the other. Open the account switcher and confirm which workspace you are in before changing anything else.

Developer Mode and Device

• 4. Developer mode is off. Adding a custom MCP server needs Developer mode. On personal accounts the toggle has been reported under Settings → Security and login → Developer mode, with app management under Settings → Plugins. Close and reopen the settings panel after toggling, because the menu often refreshes only on reopen.
• 5. You are on mobile or desktop. Custom MCP apps are created and managed on the web. Add the server at chatgpt.com first, then check the phone or desktop app, which sync afterward.

💡 Plan note: As of mid-2026, reports say full write access for custom MCP servers is a beta for Business, Enterprise and Edu plans, Pro is limited to read and fetch actions, and Free is excluded. Check OpenAI's current page before you rely on that for a purchase.

If you build integrations for a living, this is the setting that decides whether your own test server ever shows up in the list. A developer who toggles the switch on a Free account and sees nothing new has not misconfigured anything. The plan simply does not allow it.

Admin Rules and Stale Menus

• 6. Your admin turned it off. In Business, Enterprise and Edu workspaces, admins can disable apps organization-wide or require approval first. If the option is missing for you but present on a personal account, this is the likely cause.
• 7. The menu is cached. Sometimes the entry is hidden despite your eligibility. A hard refresh of chatgpt.com, fully quitting the app, or signing out and back in often clears it.

There is one more trap worth naming. In a community thread from spring 2026, admins reported that a custom connector they published did not appear in the directory for regular members. The workaround was in the Admin Console: select the connector, choose Show in directory, and share the direct connection link with the team.

A Ten Minute Fix Order

Random clicking wastes time. Run the checks below from top to bottom and stop when the connector appears.

Steps in the Right Order

1. Confirm the workspace. Switch to the account where you expect the connector.
2. Open the directory directly. Go to chatgpt.com/plugins and search for the service by name. If it lists a plugin that includes the app, that is your new entry point.
3. Look at the button. Install plugin or Connect means you are eligible. A greyed-out button means plan, region or workspace blocks it.
4. Sign in to the service. An existing connection can expire. Reconnect through OAuth and approve the permissions again.
5. Switch to the web. Mobile and desktop apps sometimes lag behind on new menus.
6. Refresh the session. Hard refresh, quit the app fully, or sign out and in.
7. Use an @ mention. After installing, call the plugin from the composer with "@" or through the More menu in the chat box.
8. Ask your admin. If you are on a team plan, send them the exact service name and ask whether it is enabled and visible in the directory.

When Nothing Works

Collect four facts before you contact support: your plan, the workspace name, the device and browser, and a screenshot of the Settings page you do see. Those four details save a round of back and forth with whoever answers. Then try a private browser window with no extensions, since an ad blocker or a stale cookie can hide parts of a settings page.

💡 Tip: Keep a short note of which connectors your team actually uses. When a menu changes, you will know in a minute whether something real broke or only moved.

Custom GPTs and Actions in the Mix

If your missing "connector" was really a custom GPT with an Action, a different change applies, and it has a deadline.

The September Retirement Notice

OpenAI announced on September 11, 2026 that custom GPTs will be retired and moved to plugins. Reported dates: new custom GPTs could no longer be created after September 25, 2026, and GPTs stop running on December 11, 2026, though timing and migration availability can vary by account and workspace. The custom GPT retirement and migration FAQ is the page to check for your own workspace.

What carries over, according to those reports:

• Instructions become a skill in the replacement plugin.
• Knowledge files are copied into plugin reference files.
• Connected apps can be carried over.
• Past conversations do not migrate, the selected model does not carry over, and migrated plugins start private, so earlier sharing settings are lost.

Actions Versus Apps

A GPT Action calls a REST API you describe with a schema. An app talks to an MCP server instead. Custom Actions may have to be rebuilt as an available app or a custom MCP server after migration, so budget time for each Action and retest your usual prompts, because migrated plugins can respond differently.

Building Your Own Connector

If no directory entry exists for the service you need, a custom MCP server is the route. It is also where most "not showing" problems for developers start.

Why HTTPS and OAuth Matter

ChatGPT connects only to publicly reachable HTTPS servers. A server on your laptop will not work without a tunnel or a hosted deployment. Add OAuth so each user signs in with their own account, and keep tool names and descriptions short and literal: the model chooses tools from those descriptions, so vague ones make the connector look broken even when it responds.

Read Versus Write Permissions

Start with read tools only: search, fetch, list. Add write tools once the read path is stable, because write actions ask for confirmation in chat and sit behind the stricter plan rules described earlier. A connector that exposes ten tools and works on three is harder to debug than one that exposes three and works on all of them.

Use GPT 5.6 Sol on PicassoIA

When the settings page defeats you, a second opinion helps, and a strong language model can turn a pile of symptoms into a clean checklist. GPT 5.6 Sol on PicassoIA is built for multi-step reasoning, accepts screenshots as input, and lets you set how hard it thinks. It cannot see your ChatGPT account, so you paste in what you see and it reasons from there.

How to run it:

1. Open the GPT 5.6 Sol page and start a new run.
2. In Image input, upload a screenshot of the Settings page you can see.
3. Write the prompt in the Prompt field, for example: "Here is my plan, my workspace type and a screenshot of Settings. Which of these seven causes fits, and what should I check first?"
4. Set Reasoning effort to medium for a quick triage, or high when the setup involves a custom MCP server and admin rules.
5. Set Verbosity to low if you only want the next action, or high if you want the reasoning.
6. If you raise reasoning effort, also raise Max completion tokens, otherwise the reply can come back empty because the reasoning used the whole budget.
7. Add a System prompt such as "You are a careful IT support engineer. Ask one question at a time" to keep answers short and practical.

💡 Alternatives: If you prefer a different style of answer, Claude Sonnet 5 and Gemini 3.5 Flash sit in the same Large Language Models category and take the same kind of pasted context.

An MCP server is also the building block behind PicassoIA's own developer API and MCP connections, which let assistants generate images and video from a prompt. Each account can run up to five predictions at the same time, shared across API and MCP use. Whether any particular assistant accepts that kind of connection depends on that assistant's own connector rules, which is exactly why the checklist above matters.

Your Turn to Create

Fixing a settings page is a chore. Making something with the result is the fun part. Every photograph in this article came from a text prompt, and you can build the same kind of visuals for your own posts, help articles and team documentation in a few minutes.

Try Seedream 5 Pro for sharp 2K images, or Qwen Image 3 for photorealistic scenes, and describe the lens, the light and the setting the way the prompts behind these photos did. When a still image is not enough, browse the text-to-video models on the all models page and bring the same scene to life.

Open Picasso IA, pick a model, and write your first prompt today. Start with a scene from your own work, such as the desk where you finally found that missing connector, and see how close the first result gets.