AI training inside ChatGPT or Claude works through an app you add to your assistant, usually called a connector. You add it once with its web address and sign in to the training service’s own account. After that, you ask for it in any chat and the lessons open right in the conversation, as an interactive screen. The app sees what you do on that screen and whatever your assistant sends to its tools. It does not get your other chats, files or connected tools on its own.
How apps inside an assistant work
Claude and ChatGPT can both connect to outside services through the Model Context Protocol (MCP), which Anthropic describes as the open standard for connecting tools to AI applications. A service that speaks MCP offers tools: named actions, each with a fixed list of what it takes and what it returns. When you ask for something, your assistant decides whether to call a tool, fills in what the tool asks for, and reads what comes back.
Some connectors go further and show a screen. Anthropic calls these interactive connectors, built on an MCP extension called MCP Apps. They appear as a card in the conversation or in a full-screen view, and the chat box stays available beside them. Anthropic’s help center says they run in sandboxed frames with strict content rules, and use the same permissions you granted when you connected the tool. A training course is this kind of connector: the lesson is the screen.
How to add one
In Claude
You add your own app as a custom connector. Anthropic’s help center lists custom connectors on Free, Pro, Max, Ultra, Team and Enterprise plans, with Free accounts limited to one. On an individual plan, go to Customize, Connectors, choose Add, then Add custom connector, and paste the service’s address. On Team plans only owners can add one, and on Enterprise, owners and people whose role allows it; they add it in organization settings, and each person then connects it from their own account.
In ChatGPT
You add your own app in developer mode, which you turn on in ChatGPT’s settings and then use to create an app from the service’s address. In a company workspace, an admin may decide who can use it; see OpenAI’s help center. Which plans include developer mode has changed over time, so check your workspace’s settings.
At work, ask your AI admin first. On company plans, an owner or admin often has to add the connector for everyone, or allow it, before anyone can connect. Anthropic also advises connecting only to services you trust and reading the permissions a connector asks for when you sign in.
What a training app can and cannot see
Apps differ, and the place to look is the list of tools an app offers and what each one takes. Your assistant can only hand an app what one of its tools asks for. In general:
- It can see what you do on its own screen (your answers, what you build), and anything your assistant puts into one of its tool calls.
- It cannot see your other conversations, your files or your other connectors by itself. Those reach an app only if your assistant sends them in a tool call that asks for them.
- Your assistant can see what the app’s tools send back to it, and the chat itself stays with your assistant’s company, under its own terms.
Skill Tree’s app, as a worked example
Here is what Skill Tree’s app offers an assistant, taken from its code:
| Tool | What the assistant sends | What it does |
|---|---|---|
| Open Skill Tree | Nothing | Opens your lessons in the chat and says what is next. Its instructions tell the assistant not to answer the exercises for you. |
| The app’s own actions | Not offered to the assistant | Carries what you do on the lesson screen. Only the app itself can call it, so moving through a lesson needs no reply from your assistant. |
| A skill was built | The skill’s id and version; for a run, yes or no, and whether it worked: yes, partly or no | Tells Skill Tree your assistant built, or ran, a skill you handed it. Never your input or its output. |
| Get the case file | Nothing | Hands your assistant a practice case to work on with you. It holds no answers. |
| Send the case results | Six short values, up to 40 characters each, and which of three pieces contain anything marked internal-only | Skill Tree compares them with its answer and replies with how many matched. No text from the pieces goes back. |
| Organization dashboard | Optionally, a team name | Only for an organization’s owners and admins. What the assistant reads is counts, not names or answers. |
Skill Tree never asks your assistant for your work, your files or your chats. Every request to the app needs a sign-in to your Skill Tree account, the lesson screen loads only Skill Tree’s own files, and when you hand a skill to your assistant, Skill Tree puts it in your message box and you press Send.
What changes when training runs in the chat
- No new tool to open. You practice in the same window where you do the work, then go straight back to it.
- What you build goes straight to use. A skill you make can be handed to the assistant beside it, instead of being downloaded and uploaded somewhere else.
- Some steps happen in your real assistant. Building a skill there, running it on your own work, or checking your assistant’s work on a case, with the app hearing back only fixed answers.