whatsapp-mcp-plus
A secure, zero-setup WhatsApp MCP server that lets AI assistants read, search, summarize, and (with user confirmation) act on personal WhatsApp chats via a safety-focused toolset, built as a maintained successor to the abandoned whatsapp-mcp.
README
whatsapp-mcp-plus
A secure, zero-setup WhatsApp MCP server. Read-first and safe by default.
It connects your personal WhatsApp to an MCP client (Claude Desktop, Cursor, Claude Code) so your AI can search, read, summarize, and, when you allow it, act on your WhatsApp, with guardrails that keep your account safe.
This is a maintained, single-binary successor to the (now abandoned)
lharries/whatsapp-mcp, rebuilt in
TypeScript on Baileys with a safety
layer it never had.
⚠️ Read the account-safety disclaimer before using. This uses an unofficial WhatsApp client. Used carelessly it can get your number banned. This tool is designed to minimize that risk, but cannot eliminate it.
WhatsApp Rewind, from your own chats
Ask your AI for whatsapp_rewind to generate a Spotify-Wrapped-style set of
story cards (1080×1920 SVG) from your history, cover, top people, your daily
rhythm, top emojis, reply-speed leaderboard, and by-the-numbers:
<p> <img src="docs/rewind/01-cover.svg" alt="Rewind cover" width="150"> <img src="docs/rewind/02-people.svg" alt="Top people" width="150"> <img src="docs/rewind/03-rhythm.svg" alt="Your rhythm" width="150"> <img src="docs/rewind/05-replies.svg" alt="Reply speed" width="150"> </p>
Prefer a single card? wrapped_card writes one shareable SVG; whatsapp_wrapped
prints a terminal card. Also: response_leaderboard (who you reply to fastest /
leave on read longest), chat_stats (per-chat, with reply times), top_words,
and export_chat.
Why this instead of the original
| original (abandoned) | whatsapp-mcp-plus | |
|---|---|---|
| Setup | Go bridge + Python server, hand-edited config, CGO on Windows | single Node process, one command |
| Safety | none (README even warns about the hole) | read-only default, allowlist, rate limits, confirm-to-send, injection guard |
| Tools | 12 (read + basic send) | 51 (react, reply, edit, delete, groups, presence, polls, transcription, analytics, chat mgmt) |
| Native deps | go-sqlite3 (needs a C compiler) |
none — uses Node's built-in node:sqlite |
| Maintained | last commit Jul 2025 | yes |
Quick start
npx whatsapp-mcp-plus
On first run it prints a QR code in your terminal. Open WhatsApp on your phone → Settings → Linked Devices → Link a device and scan it. Your recent history syncs into a local SQLite database. Nothing leaves your machine except what the AI explicitly reads through a tool call.
Or with Docker (both the connection and the server in one container):
docker build -t whatsapp-mcp-plus .
docker run -it -v wamcp-data:/app/data -v wamcp-auth:/app/auth_info whatsapp-mcp-plus
See the analytics without pairing
Want to see what whatsapp_wrapped and response_leaderboard produce? Run the
demo, it seeds a sample database and prints the real tool output:
npm run demo
Wire it into your MCP client
{
"mcpServers": {
"whatsapp": {
"command": "npx",
"args": ["-y", "whatsapp-mcp-plus"],
"env": { "WAMCP_MODE": "read-only" }
}
}
}
- Claude Desktop:
claude_desktop_config.json - Cursor:
~/.cursor/mcp.json
Safety model (safe by default)
WhatsApp bans are mostly behavioral. Reading is low-risk; cold or bulk sending is high-risk. Defaults are tuned to keep you in the safe zone.
Modes (WAMCP_MODE):
read-only(default) — every write tool is blocked. Search/read/analyze only.assisted— sending allowed, but gated by allowlist + rate limits + confirm.unrestricted— rails off (opt-in, not recommended).
Rails (on by default outside unrestricted):
- Allowlist — you can only message existing contacts or people who messaged
you first (messaging strangers is the #1 ban trigger). Override per-contact
with the
allowlist_addtool. - Rate limiting — per-minute + per-day caps and a human-like gap+jitter between sends.
- Confirm-to-send — a send returns a token and waits; nothing goes out until
you call
confirm_action. This stops a prompt-injected message from firing. - Injection guard — incoming messages that look like instructions aimed at an AI are flagged so the model treats them as data, not commands.
Every knob is an env var: WAMCP_ALLOWLIST_ONLY, WAMCP_REQUIRE_CONFIRM,
WAMCP_RATE_PER_MINUTE, WAMCP_RATE_PER_DAY, WAMCP_MIN_GAP_MS, WAMCP_JITTER_MS.
Tools
51 tools (the abandoned original had 12):
Read: search_contacts, list_chats, list_groups, get_chat,
list_messages (with date range), get_last_interaction, contact_info,
get_message_context, search_messages
Send: send_message (with reply), send_file, send_voice_note,
send_location, send_poll, send_contact
Message actions: react_to_message, edit_message, delete_message,
mark_read, forward_message, star_message
Groups: group_info, create_group, group_update_participants,
group_set_subject, group_set_description, group_invite_link, group_leave
Chat management: pin_chat, mute_chat, block_contact,
check_number_on_whatsapp, get_profile_picture
Presence / profile: send_presence, set_profile_status, set_profile_name
Media / intelligence: download_media, transcribe_voice_message
Analytics: whatsapp_rewind (6-card story set), wrapped_card (single
shareable SVG), whatsapp_wrapped (terminal card), response_leaderboard (who
you reply to fastest / leave on read longest), top_words, chat_stats
(per-chat, incl. response times), export_chat (markdown/text transcript)
Control: get_status, get_me, set_mode, confirm_action,
allowlist_add/remove/list
Voice transcription
transcribe_voice_message downloads a voice note and, if you set
WAMCP_TRANSCRIPTION_CMD (a command containing {input} that prints a
transcript), runs it through your STT of choice (e.g. whisper). Without it, the
audio is still downloaded and its path returned.
Configuration reference
| Env var | Default | Meaning |
|---|---|---|
WAMCP_MODE |
read-only |
read-only | assisted | unrestricted |
WAMCP_DATA_DIR |
./data |
SQLite DB + logs + downloaded media |
WAMCP_AUTH_DIR |
./auth_info |
WhatsApp pairing credentials |
WAMCP_ALLOWLIST_ONLY |
on (safe modes) | restrict sends to known recipients |
WAMCP_REQUIRE_CONFIRM |
on (safe modes) | two-step confirm before sending |
WAMCP_RATE_PER_MINUTE |
8 |
max sends / minute |
WAMCP_RATE_PER_DAY |
200 |
max sends / day |
WAMCP_TRANSCRIPTION_CMD |
(none) | STT command template with {input} |
WAMCP_SEND_FILE_ROOTS |
(none) | restrict which dirs files can be sent from (anti-exfiltration) |
Account safety & Terms of Service
This project uses Baileys, an unofficial WhatsApp Web client. Automating a personal WhatsApp account is against WhatsApp's Terms of Service and can result in your number being temporarily or permanently banned. To reduce that risk:
- Keep the default read-only mode unless you truly need to send.
- Never use it for cold outreach, bulk, or identical/spam messages.
- Prefer an established, well-used number over a fresh SIM; avoid VOIP numbers.
- Treat it as a personal assistant, not a marketing tool.
You accept this risk by using the tool. It is provided as-is, for personal and educational use, with no warranty. Not affiliated with WhatsApp or Meta.
Credits
Derived from whatsapp-mcp-ts
(ISC) and whatsapp-mcp (MIT). See
NOTICE. MIT licensed.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。