Let AI Use Tools

Tools let an agent retrieve information or perform an action: read a web page, check your calendar, or save an answer as a file. The model then uses the result to continue its response. You do not need to remember tool names—describe the material, task, and result you want.

Which tool do I need?

Model-driven actions need a model that actually supports tool calls. Manually converting a saved HTML file does not. Checking a capability flag in the model editor cannot add support the model lacks.

Try one small task first

1. Select a text model that supports tool calls.
2. Edit the current agent and enable the capability you need in its System section. File reading and writing have no separate system capability switch.
3. Connect an account for plugins, or grant permission when using device capabilities. An agent switch and a system permission are separate requirements.
4. Ask for one specific task, inspect the result, then continue.

Start with:

First list the writable calendars on my phone, then read tomorrow's events. Sort them by start time and name the calendar for each. Do not create or change anything.

After the result returns, replace Personal below with an actual writable calendar from the result. If none is available, set one up in the system calendar first:

In the Personal calendar you just listed, create “Organize materials” tomorrow from 3:00 to 3:30 p.m., Beijing time. Tell me whether it was saved successfully.

Separate requests help you confirm the target. To have the app request approval before eligible actions, also change the agent's approval mode to Ask when needed. A written instruction to ask first does not replace that setting.

Make requests specific

Include where the material is → the date or content range → the action → the output format → whether to make changes now.

• “Read this Feishu document: [link]. Summarize only section two as owner, task, and deadline. Mark missing details as unconfirmed. Do not create tasks yet.”
• “Turn my attached checklist into a CSV file with item, quantity, and notes columns. Save it as packing-list.csv.”
• “Use Amap to compare public transport routes from People's Square in Shanghai to Hongqiao Railway Station. Include transfers and estimated duration. Do not read my phone's location.”

One task can combine files, web pages, calendars, and plugins, but each capability must be available. Reading a document does not grant permission to create tasks; getting your location does not connect a route-planning service.

Why do only some actions ask for approval?

Automatically approve approves eligible actions, while account and system permissions still apply. New agents currently default to this mode. See Agents and Tools.

How do I know the task finished?

Tool operations appear in the response's process details. Preparing content, waiting for approval, or asking you to continue in a system app does not mean the action is complete.

• Reading succeeded: check the source and query range before relying on the summary. An empty result does not always mean no records exist.
• A file was created: open its file card or find it under Files in the sidebar. A code block or a sentence saying “created” does not establish that a file was saved.
• An external change succeeded: check the tool result and, when needed, the target calendar, document, or task.
• Only part succeeded: identify completed actions and finish only what remains.

What if it stalls or fails?

Stopping a reply, deleting a chat, or regenerating an answer does not undo completed external actions. Retrieved material may also be sent to your selected model. See Data, Privacy, and Permissions.