v1.1.0 — Now on Chrome Web Store

Your Browser. Your Sessions. Your AI.

WebMCP gives AI agents control of your actual Chrome — not a sandboxed clone, not a cloud VM. It runs entirely on localhost via the Model Context Protocol (MCP), so Claude or Cursor can navigate, click, fill forms, and read pages using the cookies and logins you already have.

100% local. Your credentials never leave your machine.

Works with Claude Desktop. Works with Cursor. Works with any MCP-compatible client.

Add to Chrome — It's Free
npm install -g webagent-mcp

Where MCP, Real Browsers, and Local Execution Meet

Every other tool gives you two of these three. WebMCP gives you all of them.

MCP-Native

Built on the Model Context Protocol from the ground up. Claude, Cursor, and any MCP-compatible agent connect directly — no adapters, no wrappers.

Real Browser

Your actual Chrome, with your cookies, your sessions, your extensions. Not a headless instance. Not a screenshot. The browser you already have open.

100% Local

No data leaves your machine. No cloud intermediary. No API keys to manage. The entire communication chain runs on localhost.

WebMCP is the only tool that is MCP-native, uses your real browser, and keeps everything local.

See It in Action

Watch an AI agent navigate, read, and interact with a web page — all through your Chrome browser.

Why WebMCP?

Human-Grade Browser Control

The agent operates through your real Chrome extension — same DOM events, same rendering pipeline, same tab you're looking at. What the AI sees is what you'd see.

Reads Pages Like an Agent Should

Every page is parsed into a structured map of every button, link, and form on the page — with numbered references. No raw HTML parsing, no fragile CSS selectors — just clean, stable targets.

Native MCP. Zero Wrappers.

Built on the Model Context Protocol from day one, not retrofitted. If your AI client speaks MCP — Claude, Cursor, or anything you build — it controls your browser with a single config line.

Fully Local. Permanently Private.

All communication stays on localhost. Your cookies, sessions, and page data never touch an external server. Password fields are automatically masked before the agent ever sees them.

25 Tools. One Protocol.

Navigate
web_navigate web_back web_forward web_refresh web_wait
Read
web_read_page web_find_element web_get_page_text web_screenshot web_highlight
Interact
web_click web_type web_select web_scroll web_hover web_press_key web_upload
Forms
web_analyze_form web_fill_form
Tabs
web_list_tabs web_new_tab web_close_tab web_switch_tab
Wait
web_wait_for_element web_wait_for_navigation
MIT licensed. Inspect every line on GitHub →

How WebMCP Compares

The browser automation space has gotten crowded. Here's where each tool actually stands.

Feature
WebMCP WebMCP
Playwright MCP Browser-Use Browserbase Stagehand
Real User Browser
MCP-Native
~ ~
100% Local / No Cloud
Uses Your Sessions & Cookies
Zero Cloud Dependency
Free & Open Source

Based on publicly available documentation as of March 2026. Corrections welcome on GitHub.

Three Steps to a Browser-Controlling AI

Install the Extension

Add WebMCP to Chrome from the Chrome Web Store (or load unpacked from GitHub). The extension opens a local WebSocket connection on port 8080 — no external network access, no sign-up.

Start the MCP Server

Run npm start in the webagent-mcp directory. Stdio mode for Claude Desktop and Cursor. SSE mode (npm run dev) for browser-based clients.

Point Your AI at It

Add one entry to your MCP client config. Your AI agent can now navigate URLs, read pages, click elements, fill forms, manage tabs, and take screenshots.

Claude Desktop Configuration
{
  "mcpServers": {
    "webmcp": {
      "command": "webagent-mcp"
    }
  }
}
macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

WORKS WITH EVERY MCP-COMPATIBLE AGENT

Claude

by Anthropic

Cursor

AI code editor

VS Code

with Copilot MCP

Any MCP Client

open protocol