Skip to main content
You create chatting agents from AI → Chatting Agents. Creating one takes a name and a choice of reply mode; everything else can be tuned afterwards.

​
Create the agent

The + New agent menu

1

Click + New agent

A menu offers three starting points.
2

Name it

Prompting Agent and Flow Agent open a short dialog with a single required Name field (for example, Sales Assistant). Click Create.Nothing is created until you confirm, so a stray click can’t leave an empty agent behind.
3

Land in the editor

A Prompting Agent opens on its General tab. A Flow Agent opens straight onto its flow canvas.

​
Starting from a template

Start from a template opens the template gallery. Filter by type — All, Voice, Chat or Assist — or type into Filter by category, then click Use template on a card. Callivox creates a new agent from that template and drops you into its editor with a reminder:
This agent was created from a template — connect your tools and knowledge base to finish setup.
The gallery has no permanent sidebar entry; reach it from the + New agent menu.

​
The General tab

The General tab of the agent editor

​
Media input

The Media input group controls what a customer may attach: “Allow customers to attach media to their messages.”
  • Accept images — the agent can look at photos and screenshots the customer sends.
  • Accept audio — “Voice notes are transcribed to text automatically.” This is always available.
  • Accept documents — the agent can read files the customer attaches.
If images or documents aren’t supported at the effort level you picked, the switch is greyed out with a note saying so.

​
Choosing an effort level

Callivox never asks you to pick a provider or a model. You pick how much thinking the agent should do, and the platform runs that tier on the best available model.

The three effort levels and their credit cost

New agents start on Medium. The footnote under the cards puts it plainly:
The platform automatically runs each tier on the best available model — no provider or model setup needed.
Effort is charged per conversation, not per message. Start on Medium, watch a week of real conversations, then move down to Low if the answers hold up. See How credits are spent.

​
Test the agent before you bind it

The Test button in the editor header runs the agent once, against one made-up message, without touching a real conversation.

The Test agent dialog

1

Click Test

The dialog reads “Send a message to this agent. Costs apply — this hits your configured LLM provider.”
2

Edit the Inbound message

It comes pre-filled with Hello, how do I reset my password?. Replace it with something a real customer would send you.
3

Click Run test

The button shows Running… while it works.
4

Read the result

You get the agent’s Reply, plus a metrics row showing how many tools it called, how many knowledge-base passages it pulled in, the tokens used, and the cost of the run. Any knowledge it retrieved is listed underneath with the source document name and a match score.
A test run is a real AI call and costs credits, the same as a live reply.
This is a single-turn test — it sends one message and shows one reply. It won’t hold a back-and-forth conversation, so use it to sanity-check tone, tools and knowledge retrieval, then bind the agent to a channel and try it for real.
Moving between tabs with unsaved edits raises Discard unsaved changes? — “You have unsaved changes. If you leave now, they will be lost.” Choose Keep editing to go back and save.

​
Next steps

Write the prompt

Give the agent its role, tone and rules — plus documents and tools.

Set up handoff

Decide when the AI hands the customer to a person.

Bind it to a channel

An agent replies only where you bind it.

Group agents into a team

Let specialists hand conversations to each other.