WhatsApp MCP
Human-in-the-loop approvals and notifications for AI agents via WhatsApp. Enables Cursor, Claude Code, and autonomous AI agents to reach users away from their computers.
README
<div align="center"> <img src="assets/logo.png" width="160" alt="WhatsApp MCP Logo"> <h1>WhatsApp MCP</h1> <p><b>Human-in-the-Loop for AI Agents via WhatsApp</b></p>
<br/>
This is an MCP (Model Context Protocol) server that enables AI agents (like Claude or Cursor) to interact directly with you via WhatsApp. It bridges the gap between your autonomous AI and your phone, allowing for runtime confirmations, permission requests, or simple status updates while you are away from your computer.
- AI agent messaging
- AI agent confirmations
- Agent interrupt system
- AI alerts
- MCP server
- Cursor MCP
- Claude Desktop MCP
- AI notifications
Perfect for
- Cursor
- Claude Desktop
- Claude Code
- OpenCode
- Codex
- Autonomous AI workflows
Example
AI: Deploy to production?
WhatsApp: [YES] [NO]
You: YES
Agent continues.
🚀 Quick Start (NPX)
Since this package is published to NPM, you can run it directly via npx in your MCP configuration.
Cursor / Claude Configuration
Add this to your MCP configuration file:
{
"mcpServers": {
"whatsapp-mcp": {
"command": "npx",
"args": [
"-y",
"@mhrj/whatsapp-mcp"
],
"env": {
"WHATSAPP_TARGET_NUMBER": "1234567890@s.whatsapp.net"
}
}
}
}
Note about Allowed Numbers:
WHATSAPP_TARGET_NUMBERis the default recipient for all messages. However, if your target is a Whatsapp Group, any member of the group would be able to answer the AI by default. To restrict replies only to specific numbers (to prevent unauthorized people from answering), you can optionally add a comma-separated list of numbers as an environment variable:"WHATSAPP_ALLOWED_NUMBERS": "123@s.whatsapp.net,456@s.whatsapp.net". If set, the AI will ONLY honor responses from these specific matching users.
Troubleshooting
npx: executable file not found in $PATHorenv: node: No such file or directory: If your IDE/Agent complains it cannot findnpxornode, it's because GUI apps (like Cursor) don't inherit your terminal's$PATH. To fix this, explicitly pass yourPATHin the MCPenvconfig."env": { "PATH": "/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin", "WHATSAPP_TARGET_NUMBER": "1234567890@s.whatsapp.net" }
Authentication (First time only)
The very first time you use the MCP server, it needs to authenticate with WhatsApp Web:
- Call the
connecttool via your Agent, or just use any WhatsApp tool normally. If the session is not authenticated yet, the server will trigger authentication automatically. - The tool result will return the QR code image directly in MCP content so modern clients can render it inline.
- If your client does not render inline images, the server also tries to write a local fallback HTML page under
~/.whatsapp-mcp/qr.html. - Open WhatsApp on your phone -> Linked Devices -> Link a Device, and scan the QR.
- The authentication session is saved to your home directory (
~/.whatsapp-mcp/baileys_auth_info), so you don't need to scan it again across restarts.
Remote "Auto-Pilot" Workflow (Skip IDE Prompts)
To make your AI agent run fully autonomously and only ask for permissions/guidance via WhatsApp (allowing you to step away from your computer):
- Remove IDE Security Blockers: When your agent (like Cursor or Claude) asks for permission to run a command or read a file, select
Always Allow(or enable "yolo" mode in settings). This gives the agent the mechanical freedom to operate without pausing indefinitely for a UI click. - Add a
.cursorrules(or system prompt) file: You must explicitly instruct the AI to use WhatsApp whenever it reaches a logical decision point. Create a.cursorrulesfile in your project root with this exact prompt:
# Human-in-the-Loop via WhatsApp
You are connected to the user via the `whatsapp-mcp` server.
Whenever you are about to make significant/destructive changes, run a deployment, or whenever you need the user's permission/opinion to proceed with a plan, you MUST use the `ask_question` tool to ask the user on WhatsApp.
DO NOT simply ask for confirmation in the IDE chat interface. You must wait for the `ask_question` tool to return the user's WhatsApp reply before proceeding with your task.
With these two steps, the AI will proactively proactively use the ask_question tool to beam its logical permission requests directly to your phone instead of freezing in the IDE.
Features & Tools
connect: Connects to the WhatsApp network. If not logged in, returns a native QR image in the tool result and may also provide a local HTML fallback.disconnect: Completely logs out of WhatsApp and invalidates the session credentials.send_message: Sends a one-way notification. Supports optional WhatsApp markdown mapping (*bold*). If the server is not connected yet, it auto-triggers authentication first.ask_question: Sends a prompt and blocks execution until a reply is received (with a timeout). Concurrent questions are smartly queued and tagged with references. If the server is not connected yet, it auto-triggers authentication first.get_incoming_messages: Polls unsolicited inbound WhatsApp messages so agents that do not support server-side sampling can still retrieve them.get_status: Provides agent connection state monitoring, including auth state such asconnected,connecting, orqr_pending.- Server instructions + prompt: The server exports built-in workflow instructions and a
whatsapp-autopilotprompt so clients can guide agents toward connect-first and WhatsApp-first behavior.
Local Development
If you'd like to run it locally from source:
- Clone the repository and
npm install npm run build- Link via absolute path instead of
npx.
License
This project is licensed under the ISC License.
Third-Party Licenses
This project utilizes the following open-source libraries:
- @modelcontextprotocol/sdk - MIT License
- @whiskeysockets/baileys - MIT License
- Other dependencies (express, pino, qrcode, etc.) are licensed under permissive licenses (MIT/BSD).
Discoverability keywords
- Human in the loop AI
- AI approvals
- Runtime approvals
- Agent notifications
- Agent approval workflow
- Cursor WhatsApp integration
- Claude WhatsApp integration
- Autonomous agents
- Cursor power users
- Remote workers
What problem does this solve?
- AI agents often get blocked waiting for approvals/questions when you step away from the IDE.
- WhatsApp MCP notifies you and collects confirmations from your phone so agents can continue safely.
Who it's for
- Cursor agent / Claude Desktop users
- Solo builders shipping agentic workflows
- DevOps & automation teams needing runtime approvals
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.