ChatGPT apps directory: five useful integrations and how to choose

How to use the directory

Open the directory, inspect an entry, and identify its publisher and included capabilities. For a plugin, follow its installation prompts, complete provider authentication when requested, and start a new chat. A provider may instead document an app connection or a separate MCP setup; use that provider’s supported connection method rather than assuming every listing has identical steps.

Do not confuse sign-in with access approval. The service account still needs the relevant permissions, and the workspace can restrict which capabilities are available. Likewise, uninstalling a plugin does not automatically disconnect its bundled connectors; remove the connection separately when that is your intention.

For a business rollout, OpenAI’s admin guidance separates plugin availability, workspace roles, allowed connector actions, service authorization, and runtime permissions. A person can see an entry yet lack permission to read a particular base or update a design.

Five options, matched to a deliverable

Canva: when the result needs to stay editable

Canva’s AI Connector guide describes creating designs, autofilling templates, finding existing designs, and exporting files through supported assistants. It currently lists eligible Canva plans rather than promising the same access to every free account. Organization controls and the user’s existing permissions also apply.

For a marketer, a useful request is a launch graphic based on an approved brief. Open the resulting design in Canva, check the copy and layout, and inspect the export at the size it will be published. The connected workflow gives the team an editable starting point; it does not approve the campaign’s claims or guarantee that a headline fits.

This is different from assuming every feature of the full Canva editor is available through the connection. Check the specific action you need, such as template autofill, rather than buying a plan based on a generic “Canva works with ChatGPT” statement.

Airtable: when records are the source of the work

Airtable describes using its ChatGPT integration to draw on bases and make updates from a conversation. Its newer MCP documentation, updated September 8, 2026, distinguishes read access from editing and says permissions reflect those granted in Airtable. Interface access and base access are not the same thing.

For example, ask for a summary of launch tasks in an authorized base. Before updating statuses from meeting notes, review the exact records and proposed field changes. Two records with similar names are not interchangeable, and an omitted field in the response does not prove the field is empty.

Airtable says MCP is included with current plans, but that does not remove the plan’s other limits or give the connection permission to every workspace. Review the granted bases and whether future bases are included in the authorization.

Figma: when a conversation should become a shared canvas

Figma’s December 2025 announcement describes FigJam diagrams, Figma Slides presentations, and Figma Buzz visual assets created from ChatGPT conversations. The workflow then moves into Figma for refinement and collaboration. The announcement includes regional and beta qualifications, so confirm current availability rather than treating its launch statement as universal access today.

A product manager could turn an approved onboarding outline into a flowchart, then ask the team to inspect the branches in FigJam. Check that each decision, label, and connection represents the intended process. A visually tidy diagram can still omit an exception or imply an incorrect dependency.

Use this for a starting artifact the team can refine. Do not infer that a generated diagram implements the workflow, updates product code, or publishes a finished design.

Photoshop: when an existing image needs adjustments

Adobe’s April 2026 Photoshop tutorial documents image editing inside ChatGPT with adjustable controls. It says the integration can be used with a free ChatGPT plan without a Photoshop subscription, and that generative edits are not supported in that documented experience.

That makes it a different job from asking an image-generation model to invent a new scene. For a supplied image, request the documented adjustment, inspect the selected area, and compare the result with the intended change. Review faces, product details, text, and the final crop before publication.

If the task requires a feature outside the connected tool, continue in an appropriate editor. Do not treat the Photoshop name as evidence that every desktop capability is exposed in chat.

Zillow: a focused example of a domain-specific app

Zillow’s description of its ChatGPT experience covers browsing actual listings and continuing to Zillow for actions such as requesting a tour. It is a useful example of a specialized connection, not a general productivity tool for every business.

A shopper can express a location, budget, and property requirements, then inspect the returned listings. Confirm the current details on the listing itself. Discovery in chat does not confirm availability, book a visit by itself, or replace the relevant professional’s advice.

Price and permission checks belong to each provider

Avoid a blanket “all apps are free” assumption. An integration may be available without a separate integration fee while its provider charges for a plan, asset, feature, or usage. OpenAI account access and provider access are separate questions.

For each connection, record who owns it, what data it may read, which changes are permitted, and where the resulting artifact or record should appear. Start with a bounded task and inspect the native result before expanding to bulk updates or customer-facing work.

Use eesel CLI to improve software-request answers

Directory discovery also creates work for IT and support: employees ask whether they can connect a newly listed app to company data. An eesel teammate can answer that question from the organization’s approved knowledge and route exceptions. A support lead can use eesel CLI directly, through a script, or through Claude Code, Codex, or Cursor to inspect how that teammate handles the request.

The CLI returns JSON and operates the same teammate and workspace as the dashboard. With Node.js 18.17+ installed, sign into the intended workspace and substitute the actual teammate name:

npx @eesel/cli status --agent "IT support"
npx @eesel/cli integrations --agent "IT support"
npx @eesel/cli instructions --agent "IT support"
npx @eesel/cli activity --agent "IT support"

Suppose the teammate tells an employee that an app is approved because it appears in the public directory. Ask the coding agent to compare the standing instructions with an approved copy of the company’s software policy. The integration list shows connections, not the contents of every policy page or a complete permission audit. If knowledge needs to be queried through teammate chat, treat that as billed live work and review the action permissions first.

The coding agent can propose a precise rule: “Treat a public directory listing as availability, not company approval. Check the approved-software source for the requested data use; route unlisted or conflicting cases to the IT owner.” The lead reviews the exact instruction diff and source reference. Inspect the update command’s --help, use --dry-run for a supported write, obtain approval, and confirm the saved configuration.

Before a test chat, check enabled actions and downstream permissions and obtain approval for the billed work. A fictional request does not make the connected teammate a sandbox. In separate fresh conversations, test an app explicitly approved for the stated use and an app that appears only in a public listing. Supply fictional request details and the relevant test-source evidence without repeating the desired rule. Check that the teammate answers the approved request and routes the other one without promising access.

Finally, inspect the reply and escalation in the actual helpdesk. This workflow changes eesel’s support behavior; it does not install ChatGPT plugins, grant provider permissions, or update the company’s approved-software register. Its benefit is a repeatable way to improve the answer from the same technical workflow the team already uses.

Choose a small set that earns its place

Start with the deliverable: an editable design, a record update, a diagram, an image adjustment, or a listing search. Confirm the provider’s current capability and your account’s access, then check the result where the work lives.

When the recurring job is answering support requests about those tools, keep the approved policy and escalation owner clear. Try eesel and use its dashboard or CLI to inspect the teammate before relying on its answers.