Tools in Chat
Web search, image generation and agent mode in the Tools menu, and how approvals work
A model does not only answer from its training. The Tools menu next to the input field gives it extra abilities: current information from the web, images, and access to your connected services and document folders. You switch a tool on, and the model reaches for it when it needs to.
The menu has two parts. Under Capabilities are three switches: Web search, Image generation and Agent mode. Below them, GPTs opens one of your GPTs in a new chat. The Tools button shows the icons of the tools that are currently on.
If the selected model cannot use tools, the switches are greyed out with the note The selected model does not support tools. Pick a different model in the header. Which tools you see at all also depends on your role.
Switch a Tool On or Off
Click a row to switch that tool on or off. Web search and Image generation are on by default, Agent mode is off. A switch you flip stays that way for your next chats too.
Your organization can set a tool to start switched off. You can still turn it on for the chat you are in, and the menu then shows This chat only. The next chat starts with it off again.
Web Search
With Web search on, the model researches live on the internet for current or factual questions instead of answering from its training. The sources it used appear as references in the answer. Use it for news, prices, releases, or anything that may have changed since the model was trained.
Image Generation
With Image generation on, ask the model for an image and it creates one right in the chat. You can refine the image further in the conversation. If more than one image model is available to you, an arrow at the end of the row opens Image model, where you pick which one to use.
Agent Mode
In Agent mode the model reads from your sources and acts on them: it searches your mail and calendar, opens files in SharePoint, reads Jira tickets or Confluence pages, and searches your document folders. It can also act, for example send an e-mail or update a ticket, after you approve it.
The arrow at the end of the row opens the list of sources: every connected integration and every document folder you can use, each with its own checkbox. The row summarizes how many are selected, for example 18 sources.
If the model cannot reach a source, check three things in this order:
- Is the service connected under Integrations?
- Is Agent mode on in the Tools menu?
- Is that source ticked in the agent mode source list?
Files you only need for one message are quicker to attach with the + button next to the input field. See Attach files.
The Sandbox
The sandbox handles common office and data formats such as Word, Excel, PowerPoint, PDF, CSV, images and archives. It runs Python and Node.js, converts documents and recognizes text in scanned PDFs. If a tool is missing, the model can usually install it. Files you attach in the chat are available in the sandbox.
- One per conversation. Each conversation has its own sandbox. No other conversation and no other user sees its files.
- Temporary. The sandbox shuts down after a while without activity, and the files in it are deleted. Results the model hands back as downloads stay with the conversation and are also listed under Generated Files. So is everything you attached. Ask for a download of anything else you want to keep.
- Limited storage. Each sandbox has a fixed amount of storage. If it reports that the disk is full, delete files you no longer need or continue in a new conversation. Other conversations are not affected.
- No access to your computer. The sandbox cannot read or write files on your device. Results arrive as downloads, and several files come as one ZIP. To let the AI write straight into a folder on your computer, use Cowork or Code in the desktop app.
- "Preparing the sandbox…" in the reply means the sandbox is still starting. The step continues on its own and you do not need to do anything. If it cannot be prepared in time, the assistant says so in its answer and carries on without it.
The sandbox can reach the internet, for example to install software packages. To block websites for yourself, set Reach websites from the sandbox to Never allow under Permissions. Installing common software packages keeps working, but a script that calls any other website fails. For web content, Web search is the more reliable route anyway.
Approve Actions
Before the model changes something in a connected service, for example sending a mail or writing to a system, the step waits for you. The answer shows Waiting for your approval, and the step names the action and its target with Deny and Approve next to it.
Click Approve to run the step once, or Deny to refuse it. The arrow next to Approve offers Allow for this chat and Always allow, so the same kind of action runs without asking again.
Once something runs without asking in the current chat, the chat header shows a button that lists it under Runs without asking. Remove an entry there to be asked again. Approvals that apply to every chat, and services you never want the model to touch, are managed under Permissions.
DeutschlandGPT's built-in features, like creating files, generating images or setting up a scheduled task, don't ask. Your role decides whether you may use them.
Artifacts and Code
Even without a dedicated tool in the menu, the model structures its outputs:
- Code blocks: Syntax-highlighted code with a copy button.
- Artifacts: Interactive outputs like charts, data visualizations, or rendered HTML appear in a dedicated panel beside the chat.
Scheduled Tasks
Ask the model in the chat to do something on a schedule, for example "every Monday at 7 a.m., summarize the industry news for me". It sets up a scheduled task, and each run sends you the result by e-mail. In the mobile app you also get a push notification.