WhatsApp MCP Server: give your AI agent real WhatsApp access

If you want an AI agent to actually use WhatsApp — read conversations, reply, send a file, check whether a number is registered — you no longer have to write the integration. WAME publishes a hosted Model Context Protocol server at https://mcp.wame.api.br. Connect your agent to it and 21 tools show up, already described, ready to call.

One command in Claude Code, or one URL pasted into Cursor or Claude Desktop. The agent discovers what it can do on its own. That is the whole setup.

What you skip

Without MCP, every capability your agent needs is a function you define, describe to the model, and keep alive as the API evolves:

const tools = [{
  name: 'send_whatsapp',
  description: 'Send a WhatsApp text message',
  parameters: {
    type: 'object',
    properties: {
      to:   { type: 'string', description: 'Number with country code' },
      text: { type: 'string' },
    },
    required: ['to', 'text'],
  },
}];
// ...plus the handler that actually calls the API, plus error handling,
// then all of it again for reading chats, sending media, creating groups...

With MCP you write none of that. The server declares the tools; the agent reads them on connect.

Connect

In Claude Code:

claude mcp add --transport http wame https://mcp.wame.api.br

In Cursor, Claude Desktop, or any client with an MCP panel: add an HTTP server and paste https://mcp.wame.api.br. The client asks for your instance authorization and that is it.

From there you just talk:

"Check whether anyone messaged sales today and never got a reply."

The agent calls find_pending_replies and list_chats by itself. You did not pick the tool — it did.

The 21 tools

They cover the whole lifecycle of a conversation, not just sending.

Talk

Read and understand

Operate

Groups

list_groups, get_group, create_group, update_group, manage_group_participants, leave_group — create a group, join, leave, add and remove participants.

Three channels, one agent

send_message takes a channel parameter. That is not an API detail — it follows from WAME delivering WhatsApp, Instagram Direct and Messenger on a single instance, behind one webhook format.

An agent that answers on WhatsApp answers on Instagram Direct without a line of extra code. Same instance, now reachable by a model.

Where this earns its place

Inbox triage. "Which conversations from yesterday went unanswered, and what was each one about?" The agent reads, summarizes and hands back a list. Nobody opens forty chats by hand.

Assisted operation. "Send the PDF receipt to the customer who asked about the invoice." It finds the contact, locates the message, sends the file.

Developer support. Inside Cursor or Claude Code with the MCP connected, you can exercise the real API while writing the code that will consume it. The agent sends a test message, reads what arrived, and you compare it against what your handler expected — without leaving the editor.

Inside an existing automation. In n8n the same MCP server plugs into an agent node as a tool. Useful when the workflow already exists and you only want to give it the ability to hold a conversation.

What MCP is not

It does not replace the API. If your system has to send an order confirmation when a status changes in your database, that is a REST call from your code, not an agent deciding at runtime. Deterministic flows still want deterministic code.

It is not a finished bot. MCP gives the agent capability. The conversation, the tone and the rules are still yours to define.

It does not widen your access. The scope is the instance you connected. And send_template — the only tool that can reach a lot of people at once — asks for confirm before it fires. That is deliberate.

Getting started

1. Create an instance in the portal. The unofficial one connects by QR code in seconds and is perfectly good for testing.
2. Connect the MCP with the command above.
3. Ask the agent something harmless: "what is my instance status?" If it answers, everything is wired.
4. Only then point it at your production instance.

The WAME MCP page carries the full tool reference, the limits on each one, and the prompts that tend to work best.

The part that compounds

Wiring AI into WhatsApp stopped being a project and became a line of configuration. The real gain is not the days saved on the integration — it is that when the API grows a new capability, the agent sees it without you rewriting anything. The new tool shows up in the list, and it gets used.

If you ship software for clients, that matters more: the same code serves every project, and now the same agent does too.