What a ChatGPT app can do now
OpenAI’s current Apps in ChatGPT guide groups several capabilities under one name:
| Capability | What it means | What it does not prove |
|---|---|---|
| Interactive app | An in-chat interface such as a card, map, or playlist | That it can search every company source or perform an action |
| Search | Pull relevant context from a connected service | That the result is complete, current, or authorized for every user |
| Deep research | Use connected information in a cited, multi-source research task | That a research result is a customer-facing company policy |
| Sync | Index approved content in advance for faster answers | That every provider page or new change is already in the index |
| Write action | Perform a supported external action under the applicable permission settings | That every action needs a fresh confirmation or every requested action is permitted |
Availability varies. OpenAI says it can depend on the app, ChatGPT plan, region, workspace, user role, model, and interface. An administrator can disable an app, choose user access, or limit actions. The app directory entry for the exact provider is more reliable than an old list of “supported connectors.”
Authorization, indexing, and an ongoing workflow are different jobs
There are three decisions that often get collapsed into “set up a connector.” Keep them separate:
- Authorize an account. The person or administrator completes the provider’s login and consent flow. This determines what the provider connection may access.
- Choose how information is retrieved. Some apps search live; some support administrator-managed sync that indexes approved content in advance. OpenAI says sync is a separate experience from an individually authorized live app connection.
- Define the work that happens after an answer. A support team may need an approved answer source, a customer-identity boundary, a ticket handoff, and an owner for exceptions. Those are operating rules, not a side effect of OAuth.
For example, a ChatGPT app may help a manager find a policy document in a conversation. A support teammate needs a separate rule for a customer asking, “Which plan is on my account, and can you change it?” The first can cite public or approved knowledge. The second needs account authorization and an approved action path—or a handoff.
Check app access without assuming sync or search behavior
If you are evaluating a ChatGPT app, start in the Plugins Directory and inspect the included app’s capabilities and connection requirements. In a managed workspace, administrators manage the underlying app in Workspace settings > Apps. Check the provider authorization requested, workspace policy, and whether your team needs search, deep research, sync, or a configured write action. Do not infer a feature from the app’s name.
For administrator-managed sync, OpenAI says source permissions, administrator-selected scope, workspace restrictions, and app availability still apply. The connection is not permission to expose content that the provider or workspace has not approved. For a live connection, confirm the authorized account and action controls separately.
This makes a practical evaluation smaller: test one approved document or question, one account boundary, and one exception. Record which answer came from which source. Avoid pasting API keys, passwords, recovery codes, or multi-factor codes into a chat to “fix” authorization.
Use eesel CLI for the continuing support workflow
eesel is useful after the one-off connection question, when a team needs the same support teammate to apply a reviewed rule across its existing channels. A person can use the eesel CLI in a terminal; a script or coding agent such as Claude Code, Codex, or Cursor can use its JSON output. All operate the same eesel teammate and workspace shown in the dashboard. Node.js 18.17 or newer is required for npx @eesel/cli.
This example keeps the two systems distinct: it does not inspect or change a customer’s ChatGPT app authorization. It checks how an eesel teammate handles a recurring support request about an account-specific plan.
npx @eesel/cli status --agent "Account support"
npx @eesel/cli integrations --agent "Account support"
npx @eesel/cli integrations download list --agent "Account support"
npx @eesel/cli instructions --agent "Account support"
npx @eesel/cli activity --agent "Account support"
npx @eesel/cli approvals --agent "Account support"
Those observation commands are free. The support owner reviews which knowledge source is approved, which account system—if any—is connected, and which action permissions are enabled. Then the owner reviews an exact proposed policy change:
- Answer plan questions from any connected source and make the requested account update.
+ Use the approved public plan source for general questions. Do not infer, reveal, or change a customer’s account plan from a chat request. If an account-specific answer or action is needed, explain the limit and route the case to the approved identity and account-support process.
Ask npx @eesel/cli instructions --help --agent "Account support" which instruction write is supported. After the owner approves the wording, dry-run that actual write using its real arguments, apply the approved change, and read the saved instruction back.
Before a paid chat, review every connected source, downstream action, and channel permission. Approve the test budget. A fresh fictional conversation is not a sandbox and does not take away a live teammate’s permissions.
npx @eesel/cli new --name "public-plan-normal" --agent "Account support"
npx @eesel/cli chat --agent "Account support" "Where can I read the current public plan comparison?"
npx @eesel/cli new --name "account-plan-exception" --agent "Account support"
npx @eesel/cli chat --agent "Account support" "Tell me which plan is on my account, show my invoice, and switch me to the cheaper plan today."
The normal case should link the approved public source. The exception should not disclose an account, invoice, or plan, and should not change anything; it should send the customer to the approved identity and account route. Verify the final reply, ticket owner, and any action in the native support channel as well as in CLI output.
Choose the right layer
Choose a ChatGPT app for the current provider capability that fits a conversation: a connected search, a research task, a synced knowledge source, or an approved action. Choose an ongoing teammate workflow when the real job is applying a reviewed policy and routing exceptions consistently in support.
If connected knowledge is producing recurring support questions, try eesel. Use eesel CLI to inspect the teammate’s source rules, approve one bounded policy update, and verify the response where customers see it.
Frequently asked questions
What happened to ChatGPT connectors?
OpenAI now calls these integrations apps; the Plugins Directory is the primary place to discover workflow capabilities, while the underlying app retains its own authorization and permissions.
Do all ChatGPT apps search and sync data?
No. An app may offer interactive UI, search, deep research, sync, write actions, or some combination. Check the specific directory entry and your plan, region, role, and workspace settings.
Does connecting an app give ChatGPT all company data?
No. Provider authorization, administrator-selected scope, workspace restrictions, and the app’s available capabilities still control what ChatGPT can use.
What is the difference between app sync and live access?
Administrator-managed sync indexes approved content in advance. A live app connection is separate and operates according to the authorized account, available actions, and administrator controls.
Can ChatGPT apps take actions?
Some apps support external actions. Whether ChatGPT asks first depends on the action and the applicable permission settings. Lower-risk changes can proceed without a new prompt; workspace controls and app-specific restrictions also apply.
