Agentic Browser MCP Server
A local MCP server that connects to a Chrome extension to enable AI agents to control browser tabs, execute automation, and interact with web pages via tools.
README
Agentic Browser MCP MVP
This repository contains a local MCP server plus a Chrome Manifest V3 extension that connects to it over a loopback WebSocket bridge.
The extension can also use Chrome Native Messaging to quietly launch a local bridge daemon and import its pairing data without opening a pairing tab.
Workspace layout
packages/server: local MCP bridge server and pairing file writerpackages/extension: Chrome extension that reconnects to the bridge and executes browser toolspackages/shared: shared protocol types and schemas
Prerequisites
- Node.js 20 or newer
- npm 10 or newer
- Google Chrome 116 or newer
Install dependencies
npm install
Build the project
npm run build
This produces:
packages/server/distpackages/extension/dist
Run the local MCP bridge server
Build and start the server:
npm run server:dev
By default the server now listens on the first free loopback port inside the shared discovery range 45320-45339, and the extension auto-scans that same range. Manual pairing is no longer required for the default setup.
When token auth is enabled, the server also opens a dedicated local pairing tab and the extension can capture the pairing information directly from that tab URL.
On startup the server still writes a pairing file to:
~/.agentic-browser-mcp/agentic-browser-mcp.pairing.json
It also logs the resolved path to stderr when the bridge starts.
Optional port and auth configuration
You can customize discovery and authentication with environment variables:
AGENTIC_BROWSER_MCP_PORT=45325
AGENTIC_BROWSER_MCP_PORT_RANGE=45320-45339
AGENTIC_BROWSER_MCP_AUTH_MODE=loopback
AGENTIC_BROWSER_MCP_AUTH_TOKEN=your-stable-token-with-at-least-32-characters
AGENTIC_BROWSER_MCP_OPEN_PAIRING_TAB=always
AGENTIC_BROWSER_MCP_PAIRING_BROWSER=\"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome\"
Notes:
AGENTIC_BROWSER_MCP_PORTpins the server to one exact port.AGENTIC_BROWSER_MCP_PORT_RANGEmakes the server try each port in order until one is free.AGENTIC_BROWSER_MCP_AUTH_MODE=loopbackenables zero-config auto-discovery.AGENTIC_BROWSER_MCP_AUTH_MODE=tokenenables stricter auth while still allowing automatic pairing through the startup tab.AGENTIC_BROWSER_MCP_AUTH_TOKENkeeps the token stable across restarts when token auth is enabled.AGENTIC_BROWSER_MCP_OPEN_PAIRING_TAB=always|neveroverrides the smart default. If unset, the tab opens automatically intokenmode and stays off inloopbackmode.AGENTIC_BROWSER_MCP_PAIRING_BROWSERlets you override how the startup pairing tab is launched. By default the server prefers Chrome and then falls back to the system browser.
URL Template Resolution
When configuring custom tools, you can use URL Templates to dynamically inject values from the current browser session or tool arguments when the tool executes. The following placeholder patterns are supported:
{{args:name}}: Resolves to the value of the custom tool argumentnameprovided by the agent.{{cookie:name}}: Resolves to the value of the cookie namednamematching the active tab's URL.{{localStorage:key}}: Resolves to the value oflocalStorage.getItem('key')evaluated in the active tab context.{{sessionStorage:key}}: Resolves to the value ofsessionStorage.getItem('key')evaluated in the active tab context.
Example
If you define a custom tool with a URL template like:
https://api.example.com/user/{{args:userId}}/profile?token={{localStorage:authToken}}&session={{cookie:session_id}}
When invoked with the argument userId: "123", the template will automatically resolve the active page's local storage and cookies to build the final request URL.
MCP Transports & Client Setup
The MCP server runs three transports concurrently in the same process:
- Stdio (Standard Input/Output)
- SSE (Server-Sent Events over HTTP)
- Streamable HTTP (MCP standard Streamable HTTP over HTTP)
By default, HTTP transports listen on port 13001. You can configure the port by setting the AGENTIC_BROWSER_MCP_HTTP_PORT environment variable.
Instructions to Connect an AI Agent / Client
1. Stdio (Claude Desktop / Cursor Configuration)
To run the server as a local subprocess, use the command entry point:
{
"mcpServers": {
"agentic-browser-mcp": {
"command": "node",
"args": ["/absolute/path/to/28_mcpsrv_in_extension/packages/server/dist/index.js"]
}
}
}
2. SSE (Server-Sent Events)
For clients that support SSE (e.g. standard remote/web clients), connect to the following endpoints:
- SSE stream endpoint:
http://localhost:13001/sse - Message POST endpoint:
http://localhost:13001/messages
3. Streamable HTTP
For standard Streamable HTTP clients, point to:
- Endpoint:
http://localhost:13001/mcp
Load the extension in Chrome
- Run
npm run build. - Open
chrome://extensions. - Enable Developer mode.
- Click
Load unpacked. - Select
packages/extension/dist.
The extension now declares the nativeMessaging permission. Once a native host manifest for com.agentic_browser_mcp.host is registered for the extension ID, the service worker will try to launch the local bridge daemon automatically on startup and on reconnect. A manifest template is included at packages/server/native-host/com.agentic_browser_mcp.host.template.json.
Manual pairing override
The extension can work without importing anything. In token mode, the startup pairing tab is usually enough. If you still want to force a specific endpoint or token manually, you can import the server-generated pairing JSON:
- Open the extension options page.
- Either:
- click the extension toolbar icon, which opens the options page, or
- open the extension details page and choose
Extension options.
- Load
~/.agentic-browser-mcp/agentic-browser-mcp.pairing.jsonwith the file picker, or paste its contents into the textarea. - Click
Import Pairing JSON.
The extension stores the validated pairing data in chrome.storage.local, prefers that endpoint first, and then falls back to range scanning if needed.
Useful scripts
npm run build
npm run extension:build
npm run server:dev
npm run server:start
npm run clean
Notes
- The extension requires the
storage,offscreen,tabs,scripting, andnativeMessagingpermissions declared in the manifest. - With the default loopback auth mode, users do not need to pair the server and extension manually.
- The native host entrypoint is exposed as
agentic-browser-mcp-native-hostand launchesbridgeDaemon.jswithAGENTIC_BROWSER_MCP_OPEN_PAIRING_TAB=neverso users are not disturbed by a startup tab. - In token mode, the extension can auto-capture pairing information from the dedicated startup tab.
- If you use token auth, disable the startup tab, and let the server generate a fresh random token on each restart, re-import the new pairing JSON through the options page.
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。