Autonomous WhatsApp agents that take action, not just answer questions.

Agent = LLM + tools + memory + goals.

A true AI agent has four properties. A language model that reasons about what to do next. Tools (typed functions) it can call to take action in the world. Memory that persists across turns and sessions so context is not lost. And goals or policies that constrain its behaviour. Without all four, you have an enhanced chatbot, not an agent. Waaru's agent layer ships all four out of the box. The LLM is your choice (Claude, GPT, Gemini, or any model with an OpenAI-compatible endpoint). Tools are pre-built for the common WhatsApp actions plus your custom REST endpoints. Memory persists per-contact across every conversation. Goals are configured per-workflow — qualify a lead, recover a cart, close a refund, escalate a complaint.

Typed tools, validated arguments, every call logged.

Every action your agent can take is a tool with a JSON-schema signature. When the agent decides to call a tool, Waaru validates the arguments against the schema before any external system is touched. If the schema is wrong, the agent gets a structured error and retries — your downstream systems never see malformed input.

• WhatsApp — send_message, send_template, send_flow, list_conversations
• Custom REST — define an endpoint with an OpenAPI spec and the agent gets the tool
• Handoff — escalate_to_human(reason, priority, summary)
• Coming soon — Shopify, Razorpay payment links, Calendar (Google / Cal.com), CRM (HubSpot, Pipedrive, Zoho), Helpdesk (Freshdesk, Zendesk), Zapier. See the integrations roadmap below.

Connect your existing AI agent to WhatsApp via Model Context Protocol.

Waaru is built MCP-native. That means more than 'we ship an MCP server' — the same operations available to humans in the dashboard are available to AI agents over the Model Context Protocol. Your existing Claude, GPT, or Gemini agent can send WhatsApp messages, look up conversation history, manage contacts, trigger flows, and read analytics, the same way it already uses any other MCP tool. This is the difference between a platform that 'has an MCP server' and a platform designed for AI agents to operate it. Waaru is the second. Your agent connects in minutes, the schema is documented, and the auth model is per-workspace API keys — no shared credentials, no impersonation risk.

Per-contact, per-workspace, persists across sessions.

Memory is the unsexy feature that decides whether an agent feels human. Waaru maintains three tiers of memory. Conversation memory — the full transcript of the active session, used as the immediate context window. Contact memory — facts the agent has learned about this specific customer across every prior conversation (preferred language, last order, complaint history, name). Workspace memory — facts about your business that every agent shares (catalog, policies, FAQs). The agent decides what to remember automatically. If a customer says 'I prefer Hindi', the agent stores that preference at contact level and switches language on the next conversation, six weeks later, on a different device.

What the agent can and cannot do, by policy.

Autonomy without guardrails is a liability. Every Waaru agent runs inside a policy envelope you configure per workspace. You set the rules for which tools the agent can call without approval, which actions require human confirmation, what the agent must refuse, and what triggers escalation. The agent obeys the policy or escalates — it never silently overrides it.

• Tool allowlists — limit the agent to a subset of tools per workflow
• Approval gates — actions over a threshold (refund > ₹5,000, manual booking change) require human confirmation
• Refusal rules — explicit topics or actions the agent will not engage on, even if asked
• Brand-voice guardrails — tone, banned phrases, mandatory disclaimers per channel
• Rate limits — per-contact and per-workspace caps on tool calls

When the agent reaches its limit, the human takes the keys with full context.

Every agent has a confidence threshold and an explicit escalate tool. When the agent escalates — or when a customer types 'speak to a human' — the conversation is routed to the team inbox with the full transcript, a one-paragraph AI summary of the situation, the contact record, and any prior notes. The human picks up at exactly the right point. On the Scale plan, the human's resolution becomes context for the agent. The next time the same intent appears, the agent handles it alone. The deflection rate compounds over months without manual retraining.

Every reply, every tool call, every escalation — logged and explainable.

You cannot trust an autonomous system you cannot inspect. Every Waaru agent conversation produces a structured trace: the prompt, the model and version that handled it, the tools considered, the tool calls made (with arguments and responses), the reply that went out, the confidence score, and the policy decisions that fired. The trace is available in the dashboard and via API. When something goes wrong — a wrong refund issued, a customer escalated to the wrong team, a hallucination — you can see exactly why, fix the policy or the tool, and re-run the failed turn on the new configuration.