Actions and tools
Every Chatbot node is an agent. It can call tools to get things done during a conversation. You control which tools it has on the Actions tab of the Chatbot node.
There are two things on this tab: your own automated actions, and the Capabilities table that governs every tool.
Automated actions
An action is a task you define for the agent, such as collecting an email for a newsletter or opening a support ticket. Each action has:
- a name,
- a description that tells the agent when to use it,
- and one or more fields to collect (text, email, number, or phone), each optionally required.
Each action you add creates a connectable output handle on the node. When the agent completes the action, the collected fields flow out of that handle into the rest of your workflow. For example, wire a “Newsletter Subscription” action to a Send Email node, or to a Post node that creates a record.
Downstream, reference a collected field as [[fieldName] from {nodeId}:{actionId}]. Every edge from the Chatbot node must connect from an action handle. See Data flow and variables.
The tools your agent can use
Below your actions, the Capabilities table lists everything the agent can do, in one place:
- Calculator and Current date and time: safe built-ins.
- Search posts and pages: searches your published content and shows the visitor rich content cards. Off by default.
- Search products (WooCommerce): searches your store and shows interactive product cards with Add to cart. Off by default, and hidden when WooCommerce is not active.
- Knowledge base search: retrieval from an attached knowledge base, when one is configured on the Model tab.
- Web search and File search: provider-native tools, available only on direct OpenAI models. These are executed by the provider.
- Your actions: each automated action appears here as a tool.
- Hand off to a human: appears when human handoff is enabled. See Human handoff.
The card tools are covered in Interactive messages.
Governing each tool
For every tool in the Capabilities table you can set three controls:
- Enabled. Turn the capability on or off. Disabled tools are never offered to the model. The system fails closed, so a tool the agent cannot see, it cannot call.
- Require approval. Pause the run and ask a human to approve before this tool runs.
- Rate limit. Cap how many times this tool may run per hour. Set 0 for unlimited.
Approval gates
When a tool requires approval, the agent stops mid-run at the point it wants to call that tool. The exact call is locked and sent to the Operator Inbox as a pending approval. An operator sees the requested action and its arguments, then approves or rejects it. Approving replays the locked call and resumes the conversation from where it paused. Use approval gates for anything that writes data or spends money.
Loop and cost guards
The agent runs a bounded tool loop. It stops when it produces an answer, or when a guard trips:
- the maximum steps per message (set under Instructions, Advanced),
- a per-run tool-execution cap,
- and loop detection, which stops the agent if it repeats the same tool call with the same arguments.
Writing tip for links
If a tool result or your instructions ask the agent to share a link, tell it to write the link in Markdown, like [View your order](https://example.com/order/123). The chat widget renders Markdown, so this makes the link clickable.