An agent can only use what you give it. Open an agent, choose Add under What it can use, and pick a kind of tool. Each tool shows a one-line summary of what it may do.

​
Tables

Pick a table and tick what the agent may do: Give the least you can. A support assistant usually needs Find and Open on orders, and Add on a notes table. Because the agent acts as the person who asked, your access rules on the table still apply. A customer’s assistant sees that customer’s orders, not everyone’s. Anything that changes data (Add, Change, Delete) follows the tool’s approval setting below.

​
Flows

Any flow can become a tool. Pick the flow and write When should it use this? in plain words:
When a customer asks for a refund
Tick This flow changes things if it writes data, sends messages or charges money, so a person is asked first. Flows are the safest way to let an agent do something with side effects: the flow does exactly what you built, nothing more.

​
Read web pages

Lets the agent open a page and read it. Leave Only these sites empty for any public site, or list the hosts you trust:
Reading outside content marks the run as having seen outside content. See Approvals and limits.

​
Search the web

Choose how it searches: the model’s own search where it has one, DuckDuckGo (no key needed), or a connected search app such as Tavily or Exa. See Connections.

​
Apps

Connect an app once under AI → Connections, then give the agent that connection as a tool. For Notion you choose what it may do (find, read, create or add to pages); for other apps you can limit it to Only these of its tools. See Connections for the full list.

​
Other agents

An agent can hand a question to another agent, for example a general assistant passing billing questions to a billing agent. Write When should it hand over? and tick That agent changes things if it does. Agents can hand over to other agents a few levels deep, and never in a loop.

​
Run code

Lets the agent run a small program to calculate something exactly instead of guessing. The code runs in the model provider’s own sandbox, not on your data or servers, so it only works with models that offer it (Anthropic and Google Gemini models).

​
Make images

Lets the agent create a picture with an image model (GPT Image, Imagen or Gemini) and give back a link to it. Choose who makes the images and which model. Pictures are saved in a bucket called agent-images, which is made for you the first time and works like any other bucket. Images are limited to 20 MB.

​
Search X

Lets the agent search posts on X. Choose an xAI model to do the searching (the agent’s own model can be any), and optionally limit it to Only these accounts. What it finds comes from outside, so it counts as outside content.

​
Find tools in big apps

Some apps, such as Zapier, bring hundreds of tools. Find tools hides them until the agent looks for the right one, so they don’t crowd every request.

​
Approval for each tool

Each tool that can change something has an approval setting:

​
Choosing tools well

  • Fewer tools make a better agent. An agent with five clear tools picks the right one more often than one with fifty.
  • Prefer a flow for anything with side effects. You decide exactly what happens.
  • Describe flows the way a person would ask for them. The description is how the agent decides when to use it.
  • Start read-only. Add change tools once you trust how it behaves.

Approvals and limits

Who decides, and how much it can spend.

Connections

Notion, GitHub, Stripe, Linear and more.