# ChatGPT apps: using tools inside ChatGPT for real work

[Canonical HTML page](https://gradiently.design/guide/chatgpt-apps)

An app turns ChatGPT from something that tells you how to do a task into something that does it, inside a real service, with your permission. Here is how the idea works and how to get finished work out of it.

## The short version

- ChatGPT apps are integrations that let ChatGPT use an outside service during a conversation, such as searching it, reading from it or creating something in it.
- OpenAI's Apps SDK for building ChatGPT apps is based on the Model Context Protocol, the open standard also used by Claude and other assistants.
- An app can only do what its connection allows, so permissions and scoped keys decide how much it can change on your behalf.
- Apps are most useful for work that ends in a real file or record, such as a design, a document or a booking, rather than a description.
- Availability of apps and custom connections depends on your ChatGPT plan, region and workspace settings, which change over time.

**ChatGPT apps** are integrations that let ChatGPT use another service while you chat: it can search that service, read your data in it, and create or change things there, after you connect it and grant permission. OpenAI's Apps SDK, which developers use to build them, is based on the Model Context Protocol (MCP), the same open standard behind Claude's connectors. The practical result is simple: instead of explaining how to make a poster or update a sheet, ChatGPT can make the poster or update the sheet, and hand you a link to check.

## Apps, GPTs, plugins and connectors

OpenAI has used several names for related ideas over the years, which causes real confusion. This is the plain difference, though menus and names do change, so OpenAI's own help pages are the place to confirm current details.

| Name | What it is | Status |
| --- | --- | --- |
| Plugins | An early way to give ChatGPT outside tools | Retired |
| GPTs | Custom versions of ChatGPT with instructions, files and optional actions | Still available |
| Connectors | Links to services such as file stores, so ChatGPT can search and read them | Folded into apps |
| Apps | Services ChatGPT can use inside a chat, sometimes with their own interface | Current approach |
| Custom MCP connections | Your own or a third party MCP server added by URL | Depends on plan and settings |

How the names relate. The common thread since 2025 is MCP, which lets one server work with many assistants.

If MCP is new to you, [what is MCP](https://gradiently.design/guide/what-is-mcp) explains it without jargon, the protocol's own documentation lives at [modelcontextprotocol.io](https://modelcontextprotocol.io), and [MCP servers for designers](https://gradiently.design/guide/mcp-servers-for-designers) shows which kinds of server are worth connecting for creative work.

## How to use apps in ChatGPT

1. **Find the app** Look in ChatGPT's apps directory or settings. Some services also give setup steps on their own site.
2. **Connect and sign in** Most apps ask you to sign in to the service and approve what ChatGPT may do there. Read the permissions before you accept.
3. **Ask in plain words** Name the app in your message, or pick it from the tools menu, then describe the outcome you want.
4. **Approve actions** For anything that creates or changes data, ChatGPT normally shows what it is about to do. Check it, then allow it.
5. **Open the result** Good apps return a link to the real thing they made, so you can review and finish it in the service itself.

## What makes an app useful for real work

Plenty of integrations only fetch information. The ones that save hours finish a job. When you try an app, look for these signs:

- **It produces a real object:** a document, a design, an entry, not just text in the chat.
- **It respects your settings:** your templates, your brand, your workspace rules.
- **The result is editable:** you can open it and change one detail without starting again.
- **Permissions are narrow:** you can let it read and create without letting it delete or spend.
- **It's easy to disconnect:** one place to revoke access when you stop using it.

## An example: designs from ChatGPT with Gradiently

Design is a good test, because an image generator can only draw a flat picture with text baked in. Gradiently works inside ChatGPT, Claude and any assistant that supports remote MCP servers. With a scoped API key, which a workspace owner or admin creates in Settings › API and agents, the assistant can search Marks, read the workspace's brand and create designs that open in the Studio. Work it does this way spends no Gradiently AI credits. Setup is in the [developer docs](https://gradiently.design/developers/mcp).

```text
Using Gradiently, make an Instagram post and a story for our
autumn tasting evening: Thursday 22 October, 7pm, tickets at
the bar. Calm, not salesy. Use our brand and our Mark.
Then give me the links.
```

A request that names the tool, the formats and the real details. The assistant fills in layout from the brand.

A portrait Instagram post for an autumn tasting evening on Thursday 22 October at 7pm

The portrait post, laid out in the brand's Mark with real, editable text.

A vertical Instagram Story for the same tasting evening, with the text centred in the frame

The story version, with the words kept in the middle band away from Instagram's interface.

Both come back as links to real designs you can adjust before exporting. [Making on brand designs from ChatGPT](https://gradiently.design/guide/design-with-chatgpt) walks through the whole flow, including what to check in your ChatGPT settings.

## Using ChatGPT apps safely

### Avoid

- Granting every permission because it's quicker
- Pasting API keys into the chat itself
- Connecting servers you found in a random list
- Letting actions run without reading them
- Leaving old connections active for months

### Do

- Allow only what the task needs
- Store keys in the client's secure settings
- Use servers published by the service itself
- Read each action before approving
- Revoke access you no longer use

An app sees what you send it and acts with the access you gave. Treat a connection like a login you lend to a capable new colleague: specific, limited and easy to take back. For a deeper look at assistants that act on their own across several steps, see [AI agents for design](https://gradiently.design/guide/ai-agent-design-workflow).

## When ChatGPT can't connect a service

Some plans and workspaces restrict custom connections, and some services need features a given client doesn't offer yet, such as sending a custom authorisation header. If that happens, the same MCP server usually works from another assistant. [Claude connectors](https://gradiently.design/guide/claude-connectors) explains the equivalent in Claude, and [connecting an AI assistant](https://gradiently.design/guide/connect-ai-assistant) covers Gradiently in several clients.

## FAQ

### What are ChatGPT apps?

Integrations that let ChatGPT use an outside service during a chat, such as searching it, reading your data or creating something in it, with your permission.

### Are ChatGPT apps the same as plugins?

No. Plugins were an earlier approach that OpenAI retired. Current apps are built with the Apps SDK, which is based on the open Model Context Protocol.

### Can I add my own MCP server to ChatGPT?

Depending on your plan and workspace settings, ChatGPT can connect custom MCP servers. Check OpenAI's current help pages, because availability changes.

### Can ChatGPT make real designs with an app?

Yes, with a design tool that offers an MCP connection. Gradiently, for example, lets ChatGPT create designs in your brand that open in its Studio for editing.