Skip to main content
Without tools an agent can only talk. With them it can look a customer up, check an order, book something, or run one of your automations.

The catalog

AI Agents → Manage Tools. Tools are grouped:

Enabling and disabling

There are two independent switches, and this is what confuses people:
1

Globally, in the catalog

AI Agents → Manage Tools — toggling a tool off removes it from every agent that used it.
2

Per agent

AI Agents → your agent → Overview → Edit → Tools — tick the ones this agent may use.
A tool enabled in the catalog is not automatically available to an agent. You must also tick it on the agent. This is the single most common reason a newly added tool is never called.

Exposing an automation as a tool

This is how you give an agent an ability you built yourself — booking an appointment, creating an order, checking stock in your own system.
1

Open the automation

Automation → your flow.
2

Turn on 'Expose as tool'

Give it a tool name (a unique identifier the AI calls, e.g. book_appointment) and a when to use it description. The description is what the model reads to decide whether to call it — write it as a plain sentence: “when the customer wants to book an appointment”.
3

Define the inputs

Each input has a name, a type (Text, Number, Yes/No), a description and a required flag. The AI fills these in, and they arrive in the run as {{trigger.<name>}}.
4

Publish the automation

An unpublished automation does not appear as a tool, however configured.
5

Enable it on the agent

AI Agents → your agent → Overview → Edit → Tools, tick it. It carries an AUTOMATION badge.
All five steps are required. Miss the publish or the per-agent tick and the tool exists but is never called.

Deleting a custom tool

Deleting removes it from the catalog and from every agent that used it. This cannot be undone. Automation tools are removed by turning off “Expose as tool” in the automation itself.

Writing a good tool description

The model chooses tools by reading their descriptions. Two rules:
  • Say when, not what. “Looks up an order by its number” is worse than “when the customer asks where their order is or gives an order number”.
  • Make inputs unambiguous. The input description is what tells the AI what to put there. customer_name with no description invites a guess.