Apps in ChatGPT reviews: what businesses should check in 2026

What changed and what to review

OpenAI now documents a shared Plugin directory for supported ChatGPT and Codex surfaces. Plugins can package skills and connectors. A connector provides access to external data or actions; installing its plugin and authorizing the external account are separate steps.

A review method that produces evidence

Start with one read-only job: find an approved policy document. Confirm the returned file, account, and citations. Then test missing content, denied access, and disconnecting the account. Only consider an action after that review identifies the identity that authorizes it and the human who owns its exception path.

This matters for support. An app can bring useful information into chat without creating the knowledge ownership, identity verification, ticket routing, action control, and handoff a production support workflow needs.

Where the product is useful—and what to examine closely

OpenAI's current plugin guide describes connections for services such as Gmail, Google Drive, and Slack. That makes a practical starting point for a team already doing research and drafting in ChatGPT: retrieve a permitted source, discuss it, and prepare an output in the same conversation. Some connectors also expose actions or interactive UI. Do not assume every integration has the same tool set.

The main operational tradeoff is that several controls govern the result. A plugin may be installed while its connector is not authorized; a connected account may lack access to the requested file; or a workspace may allow reading but disable a write. These are different causes of failure, so a useful review records the exact account, role, tool, and result—not merely whether the app was available.

Workspace controls can restrict supported connectors to read-only or an approved action set and configure when ChatGPT asks before use. External-service permissions still apply. That contradicts the blanket claim that teams have no control over apps, but it does not eliminate the need to test the intended configuration.

Also review removal and data handling. Uninstalling a plugin leaves bundled connectors connected until managed separately. Sync can index selected content in advance; non-synced use handles retrieval differently. Neither replaces the account's normal chat-retention settings or the connected service's policies.

Include the cost of completing the job

Check the current ChatGPT plan and the connected provider's terms for the exact capability you need. A seat subscription alone does not tell you the full cost of an integration rollout. Include any external-service charges, setup, review time, and maintenance. Compare those costs with the work completed, such as a verified policy lookup or reviewed draft, rather than assuming a more expensive model or plan guarantees better support outcomes.

Build a support-review record with eesel CLI

When the outcome you want is a helpdesk teammate, eesel's CLI lets you inspect and improve that teammate from your terminal or coding agent. A support lead can ask Claude Code, Codex, or Cursor to collect its instructions and connected-source information, then prepare a review of how it handles an ambiguous request. People and scripts can read the same JSON results. This is the same eesel workspace and teammate that the dashboard operates.

For a concrete review, consider cancellation wording. Use the CLI to inspect connections and instructions; the owner confirms action permissions in the dashboard and connected service, then selects a teammate appropriate for testing. A connection list is not a complete permission audit. Replace account-support below with that teammate. Node.js 18.17+ is required. Chat consumes real, billed workspace resources; the fictional example and no-action request do not restrict actual permissions.

npx @eesel/cli instructions --agent "account-support"
npx @eesel/cli integrations --agent "account-support"
npx @eesel/cli new --name "cancellation-wording-review" --agent "account-support"
npx @eesel/cli chat "Fictional review policy: explain cancellation steps, but only the account owner can request cancellation and a human handles disputed ownership. Customer says: my coworker left; cancel their account now. Draft a response and explain which fact needs verification. Do not cancel anything, send messages, change configuration, or approve actions." --agent "account-support"
npx @eesel/cli activity --agent "account-support"

Ask the coding agent to summarize the reply, whether it implied cancellation had already happened, and any action recorded in activity. That gives the owner evidence to review, rather than a vague score. If the instructions are unclear, request the precise proposed edit and its reason. The owner compares it with the actual policy and approves the change before a write; use command help and --dry-run where supported. Held-action approvals are not a universal gate on all CLI work.

Retest in fresh conversations for verified ownership, disputed ownership, and a request for a person. Finally, check actual helpdesk delivery and escalation. A CLI response is useful evidence about the teammate, but it is not by itself proof that the right human receives the case.

The verdict

Evaluate Apps in ChatGPT when your team wants connected work inside its AI conversation. Evaluate eesel for an ongoing support job with maintained knowledge and a helpdesk workflow. In either case, judge the exact task, permissions, and outcome. With eesel CLI, your existing coding agent can help prepare that evidence and make approved improvements to the same teammate your support team sees in the dashboard. Learn more or try eesel.