ChatGPT Plugins: The Plugin Directory Explained (2026)
“ChatGPT plugins” means something different in 2026 than it did in 2023, and the word came back after two years away. The original plugin store was retired. What replaced it got called connectors, then apps, and as of this July it’s called plugins again, sitting one level above apps rather than in place of them. If you searched the term expecting the 2023 store, this is what you’ll actually find.
For the full category-by-category list of what ChatGPT connects to, ChatGPT Connectors is the directory. This page is the vocabulary, the permission model, and what a plugin can and can’t do once it’s installed.
Four names, one lineage
Worth walking through, because all four are still in circulation and older guides mix them freely:
2023: plugins. A store of third-party extensions inside ChatGPT. Deprecated.
2024: connectors. OpenAI’s replacement, built around linking ChatGPT to your accounts and data rather than installing extensions.
December 2025: apps.
OpenAI renamed connectors to "apps" in December 2025. Both names are still in wide use, and nothing about how they work changed with the rename.
July 2026: plugins, again, one layer up.
Since July 9, 2026 apps ship inside "plugins," listed in one directory spanning ChatGPT and Codex. Existing connections kept working and nothing needed resubmitting; what changed is the packaging and the storefront, not the protocol.
The key sentence from OpenAI’s own documentation: “A plugin can include skills, apps, and app templates - apps remain the integrations that connect ChatGPT or Codex to external data and actions, while plugins make it easier to enable workflows in ChatGPT.”
So an app is still the thing that connects to Salesforce. A plugin is the package that ships that app alongside the skills for doing something specific with it. Existing connections were unaffected by the migration, and nothing needed resubmitting.
What a connected app can actually do
Plugins depend on the capabilities of the apps inside them. OpenAI lists five:
- Interactive apps. Rich in-chat experiences such as cards, maps, or playlists, built on the same MCP Apps extension that Anthropic and OpenAI standardized together.
- Search. Pull relevant context from a connected third-party service into the conversation.
- Deep research. Multi-source analysis with citations back to the originals. Not every app supports it.
- Sync. Index content ahead of time so responses are faster and your workspace knowledge base stays current.
- Write actions. Create or update information in a connected service. Availability depends on the app and how it’s configured.
That last one is the one people assume and shouldn’t. Write access is per-app and configurable, not a property of being connected. ChatGPT Work connectors goes deeper on where reading and acting diverge in the workspace tiers.
The permission model is the part worth reading
Apps let ChatGPT read and act in connected services. Permissions control when it asks first, and there are four settings:
| Setting | Behavior |
|---|---|
| Always ask | Asks before any app action, including reading |
| Any changes | Reads automatically, asks before changing anything |
| Important actions | Reads automatically, asks before important actions. This is the default |
| Never ask | Reads and acts automatically, with no confirmation |
“Never ask” carries elevated risk by OpenAI’s own description, and may not be available as an account-wide default even where it’s offered per app.
An important action is defined as one that could have a meaningful effect outside ChatGPT, expose sensitive information, or be hard to undo. ChatGPT evaluates the specific action and its context rather than working from a fixed list. The examples OpenAI gives:
- Sending or editing an email, message, comment, post, invitation, or appointment on your behalf
- Deleting content, cancelling an appointment, or removing a reservation
- Making a purchase, issuing a refund, or managing a transaction or subscription
- Uploading a file, or moving or renaming one in cloud storage
- Changing sharing permissions, account access, or security settings
- Sharing sensitive personal, financial, health, or identity information with an app
Context shifts the verdict. Saving a private draft is treated as lower risk than sending that same content to someone. Updating a cart is lower risk than placing the order. When one request contains several actions, ChatGPT applies the highest applicable risk level, and suspicious or hidden instructions found in app content can trigger an approval prompt or an outright block on their own.
To change the default: Settings > Apps > App Preferences > Ask permission. To override for one app, open the Plugins Directory, select the connected app, and set it in that app’s Preferences.
One thing permissions do not do: grant new access. What an app can reach was fixed when you connected it. Changing a permission only changes when ChatGPT pauses to ask before using access it already has. Removing access means disconnecting the app.
Admin controls, if you’re on a workspace plan
Workspace admins manage plugin installation under Workspace settings > Plugins, and can control each underlying app’s access and permissions either inside the plugin configuration or from Workspace settings > Apps. Admins can also configure which actions an app may take and, where available, when members must approve them.
Workspace Agents work differently: they use per-agent controls set by whoever built the agent, not the conversation-level app permissions described above.
Building your own
Custom apps are built on MCP, and the Apps SDK is OpenAI’s recommended way to package and publish one, including apps backed by MCP tools. If you already run an MCP server, ChatGPT MCP Servers covers the connection path and which plans unlock custom servers.
One wrinkle worth knowing before you plan a rollout: publishing a custom MCP app makes the app available according to your workspace settings, but it does not by itself publish a plugin. If you want it distributed as a plugin, the plugin has to be created, imported, or submitted separately.
Everything above still waits for you
Read the whole capability list again and notice what isn’t on it. Search, deep research, sync, interactive UI, write actions. Every one of them describes what happens after you type something.
These assistants can run on a schedule now — ChatGPT Scheduled Tasks and Claude Cowork both do unattended work on a timer. What none of them do is fire on an event: nothing starts because a record changed, a message landed, or a deadline passed. The schedule is the trigger, and the gap between "every hour" and "the moment it happens" is where the work falls through.
Sync gets closest, and it’s still the wrong shape: it indexes content ahead of time so your next question answers faster. It doesn’t do anything when the content changes.
So the pattern that breaks is the ordinary one. A deal closes at 4pm on a Friday and the onboarding doc doesn’t get created, the channel doesn’t get posted to, the kickoff doesn’t get booked, because nobody opened ChatGPT to ask. The connection to your CRM was there the whole time. Nothing was watching it.
Where the trigger lives
Carly is built around that specific gap. Native connections to roughly 260 apps, plus anything else with a public API through your own key on the integrations dashboard, with workflows that start on an event or a schedule instead of a prompt:
- When a deal is marked won, create the doc, post to Slack, book the kickoff.
- Every Monday at 8am, assemble last week’s numbers and send the summary.
- When an invoice goes overdue, draft the follow-up, log the task, notify the owner.
Carly is free for Zapier-style workflows, with AI agents from $35/month.
ChatGPT plugins are the right tool for working with your apps inside a conversation. Carly is the right tool when the work shouldn’t wait for you to start one.
FAQ
Are the old 2023 ChatGPT plugins coming back? No. That store was retired and its extensions were not migrated. The word returned in July 2026 for a different construct: a package containing apps, skills, and app templates, listed in a directory shared across ChatGPT and Codex.
What’s the difference between a plugin and an app in ChatGPT? An app is the integration that connects ChatGPT to an external service’s data and actions. A plugin is the package that can carry one or more apps together with skills for a specific workflow. You install a plugin; the app underneath still authenticates the same way it always did.
Do I need to reconnect my apps after the July 2026 migration? No. Existing app connections were unaffected and nothing needed resubmitting. New plugins get added from the plugin directory, and the underlying app connects and authenticates as before.
Can a ChatGPT plugin send an email or make a change without asking me? Only if you’ve set it up that way. The default permission is “Important actions,” which reads automatically but asks before anything with meaningful outside effect, including sending mail. “Never ask” removes the confirmation and carries correspondingly more risk.
Can a plugin run automatically when something happens in one of my apps? No. Plugins act inside a conversation you start. For work that begins when a record changes or on a fixed schedule, you need an automation layer such as Carly rather than a plugin.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
