Up to now, ChatGPT has known what you gave it: a message, an upload, a project, a folder on your desktop. Your actual work lives elsewhere. In your inbox, your calendar, your shared drive, your team chat, your CRM. And so far, the bridge between them has been you, copying and pasting.
Apps remove that step. Connect ChatGPT to the tools you already use and it can look things up, read the latest version, and, if you allow it, take action there. This is the lesson that turns ChatGPT from a smart assistant next to your work into one that works inside it.
What an app is
An app is a connection between ChatGPT and an outside service: Google Drive, SharePoint, Outlook, Gmail, Slack, Notion, a CRM, and many more. You'll find them under Apps in the sidebar or settings, and inside many plugins (lesson 4.2), which bundle apps with skills.
Once an app is connected, ChatGPT can use it in three ways:
- Search and read. "Find the latest version of the Acme proposal in Drive and summarise what changed."
- Bring live information into a chat. This week's calendar, the thread from yesterday, the open deals.
- Act, when you allow it. Draft and send an email, create a document, post a message, update a record.
You can let ChatGPT decide when an app helps, or call one explicitly with an @mention or from the + menu. In deep research and in Work (lesson 4.4), apps become sources and tools for the whole task.
Connecting one
Connecting takes a minute:
- Open Apps, pick the one you need, and choose Connect.
- Sign in to that service with your own work account and approve the access it asks for.
- Set its permissions (next section).
- Try it in a chat: "Using Drive, find..."
Two things to know on Business. Your admin decides which apps your workspace may use at all; if one is missing, that's a choice, not a glitch. And since late September, admins can set up some apps team-wide, so a shared tool works for everyone without each person connecting it.
Which plans: apps exist on every plan, but which apps and actions you get depends on your plan, region, and workspace. On Business many apps are on by default, and your admin can change that.
What an app can see
This is the most important rule in the lesson: an app works as you. It signs in with your account and reaches only what that account can already reach. It doesn't unlock folders you can't open or channels you aren't in.
App permissions don't grant new access. They only decide what ChatGPT may do with the access you already have.
Two more details worth knowing. An app can see the context of the chat it's used in, and your memories where relevant; ChatGPT can also save useful things it read in an app to memory. So the habits from lesson 1.4 apply to what you put in that chat. On Business, data from apps isn't used to train OpenAI's models by default. On personal plans it can be, if "Improve the model for everyone" is on.
Choosing what each app may do
For each connected app you choose how much ChatGPT may do without checking with you. There are four settings:
- Always ask. ChatGPT asks before every action in the app. The safest choice.
- Allow read actions. It can search and read freely, but asks before changing anything.
- Allow low-risk actions. Small actions, like creating a draft, go ahead; bigger ones still need you.
- Allow all actions. It acts without asking. Only offered for some apps and accounts; reserve it for apps where a mistake costs almost nothing.
Match the setting to the stakes. Reading a public knowledge base: allow read actions. Email, money, customer records, anything sent in your name: always ask. You can change the setting any time.
Using apps well
A few habits make apps work better and stay safer:
- Connect what you use, not everything. Each app's results take room in the context window (lesson 4.1), and each connection is one more way in.
- Name the source. "In Slack, #sales-team, last week" gets you a better answer than "check what people said."
- Check what it read. ChatGPT shows which app and which file or message it used. Click through when it matters.
- Read before you approve. When ChatGPT asks to send, post, or change something, check the content and the recipient. You're responsible for what goes out.
A worked example. On Monday morning: "Using Outlook and my calendar, list the meetings I have this week, and for each one find the latest related email thread and write two lines on what I need to prepare." Three tools, one answer, no copy-paste. If you ask it to email the summary to your manager, it asks first.
For the technical: custom apps
Most people never need this section. Under the hood, many apps use an open standard called MCP (Model Context Protocol), the same one Claude and others use. With developer mode, technical users can connect their own MCP app, for example to an in-house system.
Know the limits before you try. It's a beta, on the web only. Full MCP, including write actions, is for Business, Enterprise, and Edu; on Business, only admins and owners can turn it on. Pro can build apps with OpenAI's Apps SDK but doesn't get full MCP. OpenAI warns plainly that untrusted apps can expose you to prompt injection: outside content trying to give ChatGPT instructions. Write actions may need your confirmation, the riskiest ones can be blocked, and your organisation is responsible for vetting the apps it connects. If you're not the one who built the app, check with whoever did, and with IT.
That's Part 4
You can now work with ChatGPT like a power user:
- Keep the context window lean and memory accurate (4.1).
- Package repeatable jobs as skills, shipped in plugins (4.2).
- Give ongoing work a home in a project (4.3).
- Hand whole tasks to Work, on your own files (4.4).
- Connect ChatGPT to your tools, with the right permissions (4.5).
Part 5 takes it further: Codex for building and code, putting ChatGPT on a schedule so work happens without you starting it, and ready-made playbooks for common jobs.
Try it
Pick one tool you copy from most: your email, your shared drive, or your team chat. Connect it, and set its permission to Always ask to start.
Then ask a real question that needs it: "Using [app], find what we last agreed with [client or colleague] about [topic], and summarise it in three lines with links." Click through to the source to check it. Once you trust the pattern, decide whether that app deserves a looser setting, or not.
(General exercise. Role and industry versions come once accounts are in.)