Bot connectors
Workspace-scoped automated handlers that can be assigned to conversations. The dispatch path calls the bot's handler with inbound messages.
No-Code Chatbot Builder
Bots, rules, and AI — a no-code chatbot layer built on the rules engine.
Wavy does not have a visual drag-and-drop flow canvas. Instead, it combines three no-code building blocks: bot connectors that can be assigned to conversations, automation rules with trigger-filter-action logic, and AI Copilot that suggests replies and summarises conversations. Together they form a configurable chatbot layer for WhatsApp — but it is a rule-based and AI-assisted system, not a visual flow-chart builder. Bot handoff is explicit, rules respect the channel's configured state, and AI features require a server-side LLM key.
Status: Configuration-dependentBot connectors, rules engine, and Copilot paths exist. Send actions require a configured channel; AI requires an LLM key.
No visual flow canvasThe builder is rules-based (trigger-filter-action), not a drag-and-drop flow chart. Do not expect a node-and-edge editor.
Explicit bot handoffBots are assigned to conversations through connectors. They do not silently take over without configuration.
How the chatbot layer works
The three building blocks and how they combine.
This is the real bot-and-automation architecture in the backend — not a generic chatbot marketing description.
- Bot connectors — A bot connector is a workspace-scoped record that represents an automated handler. Connectors can be assigned to conversations through the inbox API. When a connector is assigned, inbound messages are dispatched to the bot's handler via
dispatch_to_bot. The bot can respond, and the conversation'shandled_byfield tracks whether a human or bot is in control. - Automation rules — The rules engine evaluates inbound messages against enabled rules. Each rule has a trigger (e.g.,
new_messagewithfirst_message_only), filters (conditions that narrow the match), and an action (e.g.,send_messagewith a body template, orunsubscribe). Rules are configured through the API or workspace UI — no code is required to create them, but they are structured as JSON trigger-filter-action objects, not visual nodes. - AI Copilot assistance — When an LLM key is configured, Copilot can suggest replies, summarise conversations, and translate messages. These are suggestions for a human agent to review and approve — not autonomous sends. Without the key, Copilot returns a not-configured response and no AI suggestions are produced.
- Human handoff — A conversation can be reassigned from a bot to a human team member at any time through the inbox API. The
handled_byfield andassigned_tofield make the current controller explicit. There is no silent bot takeover.
What you need
Requirements for a working chatbot.
Each building block has its own configuration dependency.
- Rules engine — no external provider needed — Automation rules work from the workspace database. The rules engine evaluates triggers and filters without any external API. However, send_message actions require a configured channel to deliver real replies.
- Bot connectors — require a handler — A bot connector needs a handler endpoint or function. The dispatch path calls the bot's handler with the inbound message. Without a configured handler, the connector cannot respond.
- AI Copilot — requires LLM key — Suggested replies, summaries, and translation need
LLM_API_KEYset server-side. The key points to an OpenAI-compatible endpoint (OpenAI, OpenRouter, Azure, local Ollama, etc.). Without it, Copilot is explicitly disabled. - Real message delivery — requires configured channel — All send_message actions — whether from a rule, a bot, or a manual reply — go through the WhatsAppClient. Without server credentials, sends return synthetic IDs and no real message is delivered.
Safeguards
What keeps the chatbot layer bounded.
- Explicit bot assignment — A bot does not take over a conversation automatically. A connector must be configured and assigned to the conversation. The
handled_byfield makes the current controller visible at all times. - Human handoff always available — A conversation can be reassigned from bot to human at any time. There is no lock-in that prevents a human from taking over.
- AI suggestions are not autonomous sends — Copilot produces suggestions for a human to review. It does not auto-send AI-generated replies. The suggestion-to-send step requires human action.
- Dry-run respect — Bot and rule send actions go through the WhatsAppClient. When the channel is unconfigured, sends return synthetic IDs. No real message leaves the system in dry-run mode.
- Workspace isolation — Bot connectors, rules, and conversations are workspace-scoped. A bot in one workspace cannot handle another workspace's conversations.
- AI degradation is explicit — When the LLM key is absent, Copilot returns a clear "not configured" string. It does not silently produce empty or low-quality suggestions that could be mistaken for real AI output.
Chatbot capabilities
What the three building blocks cover today.
Automation rules
Trigger-filter-action rules evaluated on inbound messages. Built-in rules handle welcome, out-of-office, and unsubscribe. Custom rules are created via API or UI.
AI reply suggestions
Copilot suggests replies for a human to review when an LLM key is configured. Suggestions are not auto-sent — they require human approval.
Conversation summaries
Copilot can summarise ongoing conversations, giving the next agent context. Requires an LLM key; degrades explicitly without one.
Human handoff
Conversations can be reassigned from bot to human at any time. The handled_by and assigned_to fields track who is in control.
Questions
No-code chatbot builder — straight answers.
Is there a visual drag-and-drop flow builder?
No. Wavy's chatbot layer is rules-based: trigger-filter-action objects configured through the API or workspace UI. There is no node-and-edge canvas. If you need a visual flow builder, that is not what Wavy provides.
Can a bot automatically take over conversations?
No. A bot connector must be configured and explicitly assigned to a conversation. The handled_by field makes the current controller visible. There is no silent bot takeover.
Do AI-suggested replies send automatically?
No. Copilot produces suggestions for a human to review and approve. The suggestion-to-send step requires human action. Copilot does not auto-send AI-generated replies.
What happens to bot replies when the channel is not configured?
Send actions go through the WhatsAppClient. When the channel is unconfigured, sends return synthetic IDs and no real message is delivered. This is the dry-run fallback.
Can I hand a conversation back to a human?
Yes, at any time. The inbox API supports reassignment from bot to human. The handled_by and assigned_to fields are updated to reflect the change. There is no lock-in.
Does the AI work without an LLM key?
No. Copilot requires LLM_API_KEY set server-side. Without it, Copilot methods return a not-configured response. No suggestions, summaries, or translations are produced. The degradation is explicit, not silent.
Build chatbots with rules, bots, and AI — all bounded by provider state.
The chatbot layer combines bot connectors, automation rules, and AI Copilot. Configure your channel for real delivery, set an LLM key for AI, and assign bots explicitly.