Knowledge lets an agent answer. A tool lets it act: look up an order in your system, hand the visitor to a booking page, or raise an event your website reacts to.
Three types, and the difference is where the work happens:
| Type | What it does |
|---|---|
| Webhook | Calls your API and uses the response |
| Client | Raises an event your own page code handles |
| Redirect | Sends the visitor to one of your pages |
Before you start
- A tool does nothing until you attach it to an agent from that agent's settings.
- For a webhook: the endpoint URL and its method. Put any API key in Secrets first — never in the URL.
Steps
- 1
Pick a tool type
The buttons are Add webhook tool, Add client tool and Add redirect tool. Pick by where the work happens, using the table above.
The three tool type buttons - 2
Set up a webhook tool
Choose the Method and the URL, then add a parameter per value the agent must send. Each parameter needs a Description — that text is what the AI reads to decide what to put in the field, so describe the value, not the field.
Never paste an API key into the URL. Create it in Secrets and reference it from the headers — the Secret dropdown lists them, and Create new secret makes one without leaving the form.
The webhook tool form - 3
Choose how each parameter is filled
Every parameter has a Value Type:
- LLM Prompt — the AI works the value out from the conversation.
- Constant Value — always the same, e.g. a store id.
- Dynamic Variable — comes from the values sent when the conversation starts.
Mark a parameter Required only if the call genuinely fails without it.
A Required parameter the agent cannot work out will stall the conversation while it keeps asking. Prefer optional with a good description.
- 4
Set up a client or redirect tool
A client tool defines the parameters your page receives, so your own code can act on them. A redirect tool is a list of name-and-URL pairs; the agent matches what the visitor asked against the name, so name your pages the way customers speak — Book a table, not reservation-v2.
A redirect tool listing pages - 5
Attach it and check usage
Back on the list, Dependent agents shows which agents use each tool, and Active / Inactive shows whether it may run at all.
After a real conversation, open it in Conversation Histories — the transcript shows the exact request, the parameters the AI extracted and the response. It is the only reliable way to debug a tool.
Good to know
- Edit as JSON is available when a form is easier to fix by hand — useful for long parameter lists.
- A tool showing errors in transcripts is the usual cause of an agent going quiet mid-conversation.
- Keep tools narrow. One tool that does one thing with two parameters is far more reliable than one that does five things.