WACrew

Guide

An AI assistant for client WhatsApp that writes like you

An AI assistant for client WhatsApp that reads groups and chats, understands voice notes and screenshots, and drafts replies in your own voice for you to approve. Open source and self-hosted.

Updated 3 min read
On this page
  1. What the assistant does
  2. You stay in charge
  3. Use it from your coding agent
  4. Install
  5. Honest limits
  6. FAQ

Most "AI for WhatsApp" products are chatbots: a menu, a flow, a polite assistant with a name. That works for a pizza shop. It does not work when the person on the other side is a client who hired you and expects to talk to you. An AI assistant for client WhatsApp has a different job: read everything, keep track of what is open, and help you answer faster without changing who is answering.

WACrew is an open-source assistant built around that idea.

What the assistant does

  • Reads every chat. Groups and direct messages, text, voice notes (transcribed), screenshots and documents. Videos can be watched with audio by a Gemini model when you ask.
  • Knows who is waiting. One queue sorted by wait time, and a summary of what each client left open.
  • Drafts in your voice. Before writing, it loads a skill measured from the messages you typed yourself: length, punctuation, abbreviations, how you open and how you disagree. You can generate that skill with the gerar_voz tool in a few seconds.
  • Reviews before sending. A deterministic check blocks the tells that give AI away, like the same opening twice, colon introductions, dashes, over-apologizing or "I hope this message finds you well".
  • Sends like a person. Read receipt first, typing indicator, the delay a human would take. A new message from the client cancels the pending reply.

You stay in charge

Every chat has a mode. Confirm is the default: the assistant shows the text and waits for your OK. Auto is per chat, for routine conversations where the answer is already decided. Even in auto, the assistant does not promise deadlines, prices or access, and it stays out of a chat as soon as you answer by hand.

If a client asks whether they are talking to a bot, the assistant does not answer and hands the chat back to you.

Use it from your coding agent

The same assistant is exposed as a WhatsApp MCP server with 37 tools, so Claude Code, Codex, Cursor or Antigravity can work the chats from where you already are. With Claude Code:

claude mcp add wacrew -e WAC_CLOUD_URL=https://your-panel.example.com \
  -e WAC_CLOUD_USER=admin -e WAC_CLOUD_PASS=change-me -- npx -y wacrew-mcp

Step by step, with the first read and the first safe reply: How to connect WhatsApp to Claude Code with MCP.

Install

git clone https://github.com/sumioshi/wacrew && cd wacrew
PANEL_USER=admin PANEL_PASS=change-me docker compose up -d

Scan the QR code at /config. The panel works on its own; the AI features need a key from the provider you choose (OpenRouter, Anthropic or any OpenAI-compatible API).

Honest limits

  • Unofficial connection. It links as an extra device through Baileys. Great for clients who already talk to you, wrong for mass messaging.
  • Your voice needs history. The voice profile is measured from what you typed. With little history it starts generic and improves as you use it.
  • It is a helper, not a decider. Anything that needs your judgment (money, scope, a complaint) goes back to you.

FAQ

Is it a chatbot? No. There are no menus or flows. It answers as you, in your chats, and you can review every message.

Can it handle more than one WhatsApp number? Yes. One MCP server can talk to several panels, and every tool takes a conta parameter to pick the account.

Where does my data go? To your own server. WACrew has no hosted service; only the AI calls go to the provider you configure.

Does it understand voice notes in Portuguese and English? Yes. Transcription runs locally on a Mac or through an API, and the text becomes searchable.