# Use Wundaly with your AI

> Connect Claude, ChatGPT, Gemini, Grok or any MCP assistant to Wundaly: it works inside your apps, remembers, and waits for your OK before anything risky goes out.

Source: https://wundaly.geenius.io/docs/use-with-your-ai/

Wundaly has no AI of its own inside. Your own AI, Claude, ChatGPT, Gemini, Grok or any assistant that supports MCP, connects once and then works in your apps: projects and tasks, documents and tables, invoices and quotes, calendar and bookings, contacts, forms, files, websites and more.

> [!HONEST]
> Your AI can do a lot, and you stay in charge. Emails to customers and other risky steps wait in Approvals until you say OK, each agent has only the permissions you give it, and every step is written in the audit trail. You never paste a key or a password into a chat.

## Connect it

Open the connector settings of your AI, add Wundaly with the connector address from your invitation, and approve it once. It then works in every new chat. See [connect your AI](/docs/connect-your-ai/).

## How your AI sees it

Behind the connector your AI gets one tool per app. Each tool takes a `tool` (the part of the app) and an `action` (the verb). You never type these: your AI does.

```json title="What your AI sends"
{ "tool": "approval", "action": "new", "args": {} }
```

## Your first ask

```text title="A first ask"
Read the project "Studio Lina" and tell me where we are. Then draft the invoice for the website work. Do not send anything.
```

Your AI reads the project, drafts the invoice and puts the email in Approvals. You open it, edit it and approve.

## The tools

::mcp-tools

Ask for the actions of a tool with `action: "help"`, and for one action's arguments with `action: "help"` and `of`. There is one more tool for each app you install, such as geenius_mvp or geenius_video.

## Habits that work

- **Name the project once.** "This is for Studio Lina". Your AI keeps it in mind from the workspace, not from the chat.
- **Ask it to read first.** "Read the project and tell me where we are."
- **Check the approval.** Open it, read exactly what will happen, then approve.
- **Give each agent a job.** One for the website, one for the books.

## What your AI cannot do

It cannot approve its own work: only a person decides an approval. It cannot do what you did not give its agent permission for, and it cannot hide a step: the audit trail keeps them all.