Both projects connect your own WhatsApp number to an AI agent as a linked device and expose it over MCP. They were built for different jobs. lharries/whatsapp-mcp is a lean bridge: search your chats, send a message or a file. WACrew is an app for answering clients, with an MCP server on top. This page puts them side by side using each README, checked on November 14, 2026.
At a glance
| lharries/whatsapp-mcp | WACrew | |
|---|---|---|
| Connection | whatsmeow (Go), WhatsApp Web multi-device | Baileys (TypeScript), WhatsApp Web multi-device |
| Pieces | Go bridge + Python MCP server (uv) | Collector + Next.js panel + MCP server (wacrew-mcp on npm) |
| Storage | SQLite in whatsapp-bridge/store/ | JSONL and media per chat in data/ |
| MCP tools | 12 | 36 |
| Voice notes | Download the file; send voice notes (FFmpeg converts to Opus) | Transcribed with Whisper (MLX on Apple Silicon or via API); send voice notes |
| Sending | Direct send_message / send_file | Ticketed delivery with read receipt, typing and pauses; review before sending |
| Human approval | Up to your client | Confirm mode per chat, on by default |
| Runs | Locally | Locally or 24/7 with Docker, Railway template |
| License | MIT | Apache 2.0 |
| Last commit | July 2025 | November 2026 |
Setup effort
lharries/whatsapp-mcp needs Go, Python 3.6+, uv and optionally FFmpeg. You run go run main.go, scan the QR code, then paste a JSON block with absolute paths to uv and the server folder into Claude Desktop or Cursor. On Windows you also need CGO and a C compiler for go-sqlite3. The README notes you may need to re-authenticate after about 20 days.
WACrew needs Node 20+ and FFmpeg, or just Docker. On a server: docker compose up -d, open /config, scan the QR code. Then one line in your client:
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
What the agent can do
lharries/whatsapp-mcp covers the essentials well: search_contacts, list_messages, list_chats, get_chat, get_message_context, send_message, send_file, send_audio_message, download_media and a few contact helpers.
WACrew adds tools for the work around a reply: caixa_de_entrada (who is waiting, across every chat), transcrever (voice notes), ver_imagem and assistir_video (screenshots and screen recordings), ler_documento (PDF, docx, csv), preparar_resposta and revisar_texto (your style and a pre-send review), status_envio and cancelar_envio (delivery tickets), CRM notes, tags and Google Calendar. Full list in the 36 tools. Tool names are in Portuguese; descriptions guide any model.
Sending: the real difference
With lharries/whatsapp-mcp, send_message sends immediately. That is the right behavior for a bridge, and your MCP client decides whether to ask first.
WACrew treats sending as the risky step. responder returns a ticket; the message goes out later with a read receipt, typing indicator and pauses, and the delivery stops if the contact writes again or you reply from your phone. Before that, more than 25 code rules block robotic tells and repeated bubbles. In Confirm mode, the default, the agent has to show you the text first.
Which one to pick
- Personal use, quick experiments, you are fine with Go and Python: lharries/whatsapp-mcp (or a maintained fork, see alternatives).
- You answer clients, often in groups, and want an inbox, transcription and replies that read like you: WACrew.
- You only need read access: neither; look at a read-only server in best WhatsApp MCP servers.
FAQ
Can I run both on the same number? Each links as a separate device, and WhatsApp allows several linked devices, so yes. Two agents able to send to the same chats is a recipe for double replies, though.
Does WACrew read my old history? No. Capture starts when you pair the number. lharries/whatsapp-mcp stores what the phone syncs to the bridge after linking.
Which is safer? Both use an unofficial client, so the account risk is similar and depends on behavior. WACrew's confirm mode and human-paced delivery reduce the chance of an agent sending something you did not intend.



