Rule-Based Chatbots

Rule-Based Chatbots are deterministic, token-free workflows. They do not invoke an LLM. Each incoming message follows this model:

Trigger -> Action -> Success or Failure -> Response or Next Node

Enable the mode

Open an agent in Studio.
Open Overview.
Set Conversation mode to Rule-Based Chatbot.
Open Rule Flow.
Create and save the nodes and actions.

Rule bots use the existing messaging pipeline, channel adapters, MCP connections, and HTTP connections.

Nodes

A node is one deterministic step in the conversation.

Field	Meaning
Title	Name shown in the builder.
Trigger type	How the node is selected.
Trigger value	Command, keyword, exact text, or interaction value to match.
Response	Text returned after the node succeeds.
Failure response	Safe text returned when an action fails.
Enabled	Whether this node can be matched.

Supported trigger types:

slash_command: /start, /help, or another command.
button_click: a button or interaction payload.
menu_selection: a platform menu selection.
keyword: a supported keyword/text match.
text: an exact text match.

Actions and routes

Actions are interactions attached to a node. The builder’s Save button persists the node and action fields. If the node has no action yet, Save creates one when action settings have been entered.

Field	Meaning
Action type	Response, route, MCP action, or HTTP action.
Action label	Visible button/menu text.
Action payload	Stable value sent back when selected.
On success, continue to	Target node. Empty means the flow ends.
Failure response	User-facing fallback for failed actions.

Action types:

Response only: return the node response and end.
Route to another node: continue to the selected node.
Route by slash command, keyword, or menu selection: use the configured payload as the next route.
Execute MCP tool: run an existing MCP tool attached to the agent.
Execute HTTP tool: run an existing HTTP API tool attached to the agent.
Execute action + continue: run a tool, then route on success.

The action payload is the routing value. Keep it stable even if the visible label changes. Create a node:

Title: Start
Trigger type: Slash command
Trigger value: /start
Response: Welcome. Choose an option below.

Add two actions:

Label: View services
Payload: services
On success: Services

Label: Contact support
Payload: support
On success: Support

Create Services and Support nodes with their own responses.

/start -> Start
          |-> View services -> Services
          |-> Contact support -> Support

Recipe: HTTP lookup

Configure an action with:

Action type: Execute HTTP tool
Action label: Check status
Action payload: check_status
On success: Status result
Failure response: We could not retrieve the status right now.

Use workflow variables in the tool arguments:

{
  "variables": {
    "reference": "{{workflow.reference}}"
  }
}

Configure the result node response:

Current status: {{tool_result}}

Recipe: MCP action

Attach an MCP connection to the agent, then configure:

Action type: Execute MCP tool
Action label: Run verification
Action payload: verify
On success: Verification result
Failure response: Verification is temporarily unavailable.

The rule engine passes the configured arguments to the existing MCP executor. Domain-specific interpretation belongs in the tool schema, not in the rule engine.

Recipe: multi-step flow

Use one node for each step:

/start
  -> Collect name
  -> Collect email
  -> Confirm details
  -> Execute action
  -> Complete

Use workflow state and templates to carry values forward. Use /cancel to reset a conversation during a multi-step flow.

Templates

Responses can reference the latest action result:

The result is: {{tool_result}}

Tool arguments can reference workflow state:

{
  "arguments": {
    "reference": "{{workflow.reference}}"
  }
}

Platform behavior

The backend emits platform-neutral interactions. Adapters translate them into each channel’s supported markup, including Web Chat, WhatsApp, Telegram, Slack, and Discord. The same label and payload are therefore reusable across channels.

Testing checklist

Test all of the following before deployment:

/start or the first entry trigger
Every button payload
Response-only actions
Successful MCP and HTTP actions
Failed actions and configured fallback text
Success routes to another node
A node with no target, which should end the flow
/cancel during a multi-step flow
Agent Preview and Web Chat
Native button rendering on every connected platform

Troubleshooting

The node does not match

Check that the node is enabled and that the trigger type and value match the incoming text or payload.

A button appears but does not route

Check that its action payload is saved and that the target node belongs to the same agent.

The action succeeds but the response is empty

Check the target node response and template names. Use {{tool_result}} for the latest action result.

The tool fails

Configure a failure response on the action. Internal executor errors should not be exposed to users.

No LLM response appears

That is expected. Rule-Based Chatbots intentionally bypass the LLM pipeline.

Get Started

Enable the mode

Nodes

Actions and routes

Recipe: welcome menu

Recipe: HTTP lookup

Recipe: MCP action

Recipe: multi-step flow

Templates

Platform behavior

Testing checklist

Troubleshooting

The node does not match

A button appears but does not route

The action succeeds but the response is empty

The tool fails

No LLM response appears

​Enable the mode

​Nodes

​Actions and routes

​Recipe: welcome menu

​Recipe: HTTP lookup

​Recipe: MCP action

​Recipe: multi-step flow

​Templates

​Platform behavior

​Testing checklist

​Troubleshooting

​The node does not match

​A button appears but does not route

​The action succeeds but the response is empty

​The tool fails

​No LLM response appears

Enable the mode

Nodes

Actions and routes

Recipe: welcome menu

Recipe: HTTP lookup

Recipe: MCP action

Recipe: multi-step flow

Templates

Platform behavior

Testing checklist

Troubleshooting

The node does not match

A button appears but does not route

The action succeeds but the response is empty

The tool fails

No LLM response appears