WebMCP
Enables websites to expose JavaScript functions as MCP tools, allowing AI agents to interact with the browser environment via a protocol and Chrome extension.
README
[!IMPORTANT] 🚨 The WebMCP API has been accepted as a W3C deliverable and is transitioning to official web standard development. The community and all related work has moved to the WebMCP explainer repository. Please join us there! 🚨
[!IMPORTANT] 📝 Repository Status: The MCP-B extension is no longer open source. This repository represents an older version of the codebase and is maintained for historical reference only. All open source content (transports, etc.) have been ported to the WebMCP Github Org.
WebMCP:
🚀 Quick Start • ✨ Live Demo • 📚 Documentation • 🤝 Contributing
Join Our Community

The protocol
WebMCP is the underlying protocol which MCP-B implements. It is a protocol which exposes function in your browser javascript to LLM's as MCP tools.
For a more indepth understanding, refer here: https://mcp-b.ai/blogs
Live Demo
See MCP-B in action right away:
Examples
Check out our examples repository: MCP-B Examples
The examples repository contains:
- Vanilla TypeScript Demo: A simple todo app where MCP tools allow AI to manage tasks
- React Demo: Modern React application with MCP-B integration
- Script Tag Demo: The simplest integration - add MCP-B to any website using just a script tag
Community Examples
These demos highlight how MCP-B integrates into websites without needing complex setups. Install the MCP-B Chrome Extension to interact with the tools via the extension's chat interface or tool inspector.
What is MCP-B?
MCP-B extends the Model Context Protocol (MCP) with browser-specific transports, allowing your website to act as an MCP server. Websites expose existing functionality (e.g., APIs, forms, or state) as structured tools that AI agents can call directly.
Key components:
- Tab Transports: Use
postMessagefor communication between your website's MCP server and clients in the same tab. - Extension Transports: Use Chrome's runtime messaging for communication with browser extensions.
This setup enables AI to interact with your site deterministically, respecting user authentication (e.g., session cookies) and scoping tools to specific pages or user states.
Quick Start
Get MCP-B running on your website in minutes. This guide focuses on adding an MCP server to expose tools, using the examples as a blueprint.
Prerequisites
- Node.js 22.12+ (check with
node --version) - pnpm 10+ (install via
npm install -g pnpm) - A website with JavaScript (vanilla, React, etc.)
- MCP-B Chrome Extension installed for testing
Development Setup (Repository Contributors)
If you want to contribute to MCP-B or run the examples locally:
-
Clone and install:
git clone https://github.com/MiguelsPizza/WebMCP.git cd WebMCP pnpm install pnpm build:shared # Build internal shared packagesNote: Some postinstall scripts may fail initially - this is normal.
-
Configure your development extension ID (optional):
# If your extension ID differs from the default cp apps/native-server/.env.example apps/native-server/.env # Edit apps/native-server/.env with your extension ID -
Configure extension model provider and model name:
# Create a .dev.vars file in apps/backend from the example cp apps/backend/.dev.vars.example apps/backend/.dev.vars # Edit apps/extension/.env with Open AI or Anthropic API Keys # ANTHROPIC_API_KEY="your_claude_api_key_here" -
Start development:
pnpm devThis automatically:
- Builds all packages and native server
- Registers native messaging host for both production and dev extension IDs
- Starts WXT with persistent browser profile
- Launches extension in Chrome with hot reload
- Starts documentation website and all package watchers
-
Find your extension ID (if needed):
- Open Chrome at
chrome://extensions/ - Enable "Developer mode"
- Find your MCP-B extension and copy the ID
- Update
apps/native-server/.envwithDEV_EXTENSION_ID=your-extension-id - Restart
pnpm dev
- Open Chrome at
-
Run examples - See the MCP-B Examples Repository for example applications.
Adding MCP-B to Your Existing Website
For adding MCP-B to your own project (recommended for most users):
Step 1: Install Dependencies
npm install @mcp-b/transports @modelcontextprotocol/sdk zod
Step 2: Add an MCP Server to Your Website
Create a single MCP server instance and connect it via Tab Transport. Expose tools that wrap your existing logic.
Example (vanilla JS/TypeScript):
import { TabServerTransport } from "@mcp-b/transports";
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { z } from "zod";
// Create the server (one per site)
const server = new McpServer({
name: "my-website",
version: "1.0.0",
});
// Expose a tool (wrap your app's logic)
server.tool("getPageInfo", "Get current page info", {}, async () => {
return {
content: [
{
type: "text",
text: JSON.stringify({
title: document.title,
url: window.location.href,
}),
},
],
};
});
// Connect the transport
await server.connect(new TabServerTransport({ allowedOrigins: ["*"] })); // Adjust origins for security
- What this does: The server listens for clients (e.g., the extension injects one). Tools like
getPageInfobecome callable by AI. - Tips: Use Zod for input schemas. Add visual feedback (e.g., notifications) so users see AI actions.
Step 3: Test It
- Run your site (e.g., via a dev server).
- Visit the page in Chrome with the MCP-B extension installed.
- Open the extension popup:
- Go to the "Tools" tab to see your exposed tools.
- Use the chat interface to ask AI to call them (e.g., "Get the page info").
- Or manually invoke via the inspector.
Step 4: Explore Examples
Check out the MCP-B Examples Repository for ready-to-run starters:
- vanilla-ts: Basic todo app. Tools:
createTodo,getTodos, etc. Demonstrates dynamic tool registration and UI updates. - react: Modern React application with hooks and state management integration.
- script-tag: Simple MCP-B integration using just a script tag - no build tools required.
Copy patterns from these examples to your site. Focus on wrapping client-side functions—e.g., use fetch with credentials: 'same-origin' for authenticated calls.
For more, see the documentation.
Using the Extension
The MCP-B extension acts as a client that discovers and routes calls to your website's MCP tools. Users can interact via a chat interface or tool inspector.
Option 1: Download from Chrome Web Store
Install the official release: MCP-B Extension.
Once installed:
- Visit your MCP-enabled website.
- Open the extension popup (click the icon in the toolbar).
- Use the chat to query AI (e.g., "Add item to cart") or the inspector to list/call tools manually.
Option 2: Build from Source (Dev Build)
For the latest features or custom modifications:
- Clone the repo:
git clone https://github.com/MiguelsPizza/WebMCP.git. - Install:
cd WebMCP && pnpm install && pnpm build:shared. - Build the extension:
pnpm --filter @mcp-b/extension build. - Load in Chrome: Go to
chrome://extensions/, enable Developer Mode, and load./apps/extension/.output/chrome-mv3unpacked.
Run in dev mode for hot reloading: pnpm --filter @mcp-b/extension dev.
Hooking Up the Native Server
To connect MCP-B to local MCP clients (e.g., Claude Desktop, Cursor) via a native server, bridging the browser to local processes:
IMPORTANT: You will need to disable the chrome webstore version of the extension if you have it downloaded. Failure to do so will cause port clashing when the dev and prod extension run
- Install globally:
npm install -g @mcp-b/native-server. - Run the host:
@mcp-b/native-server(starts a server on port 12306 by default).
Add this configuration to your MCP client (e.g., in Claude's config or Cursor's .cursor/mcp.json):
{
"type": "streamable-http",
"url": "http://127.0.0.1:12306/mcp",
"note": "For Streamable HTTP connections, add this URL directly in your MCP Client"
}
- What this does: The native server proxies requests from local clients to the browser extension, allowing tools from your website to be called from desktop apps.
- Note: The native server is based on mcp-chrome by hangwin. Ensure the extension is running and tabs with your site are open.
Test by running a local client (e.g., MCP Inspector) pointed at the URL, then calling tools from your site.
Advanced Usage
- Dynamic Tools: Register/unregister tools based on page or user state (e.g., admin-only tools in React components).
- Tool Caching: Annotate tools with
{ annotations: { cache: true } }to persist across tabs. - Security: Tools run in your page's context—only expose what you'd allow via UI. Use MCP's elicitation for sensitive ops (support coming soon).
Repository Structure
WebMCP/
├── apps/ # Application packages
│ ├── extension/ # Browser extension
│ ├── backend/ # Backend server (Cloudflare Workers)
│ └── native-server/ # Native messaging host
├── shared/ # Internal shared packages
│ └── utils/ # Shared utility functions
└── e2e-tests/ # End-to-end tests
Related Repositories
- NPM Packages - Core npm packages (@mcp-b/transports, @mcp-b/mcp-react-hooks, etc.)
- Examples - Starter projects and demos
- Web Demo - Full-stack demo site with documentation
- WebMCP Userscripts - Tampermonkey scripts that inject MCP-B servers into popular websites
Troubleshooting
Common Setup Issues
Git clone times out:
# If the initial clone fails, complete it manually:
git clone https://github.com/MiguelsPizza/WebMCP.git
cd WebMCP
git pull origin main
Native server postinstall errors:
# These errors during pnpm install are normal and can be ignored:
# "Cannot find module '/path/to/apps/native-server/dist/scripts/postinstall.js'"
# The packages will still build correctly.
Example won't start: See the MCP-B Examples Repository for proper setup instructions.
Import resolution errors:
For monorepo development:
# Ensure the workspace is properly built:
pnpm build:shared # Build internal shared packages
pnpm build:apps # Build all applications
# Or run from the root with workspace support:
pnpm dev
For examples: See the MCP-B Examples Repository.
Port conflicts:
- Main dev server runs on port 5173-5174
- Extension dev server runs on port 3000
- Native host runs on port 12306
Extension Issues
Extension not detecting tools:
- Ensure the extension is installed and enabled
- Refresh the page after starting your MCP server
- Check the extension popup "Tools" tab
- Look for console errors in browser DevTools
Tools not working:
- Verify your
TabServerTransportconfiguration - Check that
allowedOriginsincludes your domain - Ensure tools are properly registered before transport connection
Development
git clone https://github.com/MiguelsPizza/WebMCP.git
cd WebMCP
pnpm install
pnpm build:shared # Build internal shared packages first
pnpm dev # Runs all in dev mode
Contributing
Contributions welcome! Focus on transports, examples, or docs. See CONTRIBUTING.md.
Security & Trust
- Respects browser sandbox and same-origin policy.
- No data collection; runs locally.
- Audit tool calls via the extension.
Roadmap
- Firefox/Safari support.
- Full MCP spec (beyond tools).
- Native host upstreaming.
Tutorial video
MCP-B lets your website become an MCP server, exposing functionality as tools that AI agents can call directly—using the browser's existing authentication and security model.
License
AGPL-3.0 - see LICENSE.
Created by Alexander Nahas (@miguelsPizza). Reach out: alexnahasdev@gmail.com.
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。