Browser Mcp Lite

browser-mcp-lite

[](https://www.npmjs.com/package/browser-mcp-lite) [](https://opensource.org/licenses/MIT) [](https://modelcontextprotocol.io) [](https://developer.chrome.com/docs/extensions/) [](https://github.com/notoriouslab/browser-mcp-lite)

[繁體中文](README.zh-TW.md) | English

Give your AI assistant eyes into your real browser — in ~500 lines of code.

A minimal, auth-secured MCP server that lets AI assistants read pages, take screenshots, and run scripts in your actual Chrome browser — with your existing login sessions intact.

AI Assistant ──HTTP POST──▶ MCP Server (:12307) ──WebSocket──▶ Chrome Extension ──▶ DOM
                            (token auth)           (ws://)       (minimal permissions)

Why build your own?

Existing browser MCP solutions are either:

• Too heavy — mcp-chrome requires debugger, history, `` permissions, ships minified code you can't audit
• Headless only — Playwright MCP launches a new browser, can't see your logged-in sessions
• Cloud-dependent — Browserbase, etc. route your pages through third-party servers

browser-mcp-lite takes a different approach:

| | mcp-chrome | Playwright MCP | browser-mcp-lite | |---|---|---|---| | Reads your real browser | Yes | No (headless) | Yes | | Auth on MCP endpoint | No | No | Token auth | | Extension permissions | 8+ | N/A | 5 (minimal) | | Code you can audit | Minified | Yes | ~500 lines | | Login sessions | Yes | No | Yes |

What it does

Four tools, exposed via standard MCP protocol:

| Tool | Description | Returns | |------|-------------|---------| | list_tabs | List all open tabs | [{id, url, title, active}] | | read_page | Read page as accessibility tree | Structured text (token-efficient) | | screenshot | Capture visible area | Base64 PNG image | | inject_script | Run custom JS in a tab | Script return value |

Quick start

1. Install and start the server

Option A — npx (no clone needed):

npx browser-mcp-lite

Option B — clone:

git clone https://github.com/notoriouslab/browser-mcp-lite.git
cd browser-mcp-lite/server
npm install
node index.js

On first run, a token is auto-generated. The server prints the full token and a ready-to-paste config block:

[browser-mcp-lite] MCP server: http://127.0.0.1:12307/mcp
[browser-mcp-lite] WebSocket:  ws://127.0.0.1:12307/ws

⚠ First run — new token generated

━━━ Auth Token (paste into Chrome Extension popup) ━━━

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

━━━ MCP Client Config (save as .mcp.json) ━━━
{
  "mcpServers": {
    "browser": {
      "type": "http",
      "url": "http://127.0.0.1:12307/mcp",
      "headers": {
        "Authorization": "Bearer "
      }
    }
  }
}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

[browser-mcp-lite] Waiting for Chrome Extension...

2. Load the Chrome Extension

1. Open chrome://extensions/
2. Enable Developer mode (top right)
3. Click Load unpacked → select the extension/ folder
4. Click the Browser MCP Lite icon in your toolbar
5. Copy the token printed in the terminal and paste it into the Auth Token field
6. Click Connect

The extension auto-connects. Click the toolbar icon to verify the green dot.

3. Connect your AI assistant

> Note: The ready-to-paste .mcp.json config is printed in the terminal when the server starts.

Claude Code

Add to your project's .mcp.json:

{
  "mcpServers": {
    "browser": {
      "type": "http",
      "url": "http://127.0.0.1:12307/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN_HERE"
      }
    }
  }
}

Replace YOUR_TOKEN_HERE with the token printed in the terminal.

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "browser": {
      "type": "http",
      "url": "http://127.0.0.1:12307/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN_HERE"
      }
    }
  }
}

Cursor / VS Code

In Cursor: Settings → MCP → Add Server:

• Name: browser
• Type: http
• URL: http://127.0.0.1:12307/mcp
• Headers: Authorization: Bearer YOUR_TOKEN_HERE

Architecture

Three components

┌─────────────────┐     ┌──────────────────┐     ┌──────────────────┐
│   AI Assistant   │     │    MCP Server    │     │ Chrome Extension │
│ (any MCP client) │────▶│   (Fastify)      │────▶│   (Manifest V3)  │
│                  │HTTP │ :12307/mcp       │ WS  │                  │
│                  │POST │ Token auth       │     │ tabs, scripting,  │
│                  │◀────│                  │◀────│ activeTab, alarms │
└─────────────────┘     └──────────────────┘     └──────────────────┘

MCP Server (server/index.js, ~150 lines)

• Fastify HTTP server with @modelcontextprotocol/sdk Streamable HTTP transport
• Token auth on every request (Bearer token from ~/.browser-mcp-secrets.json)
• WebSocket endpoint for Chrome Extension communication
• Per-session MCP transport management

Chrome Extension (extension/background.js, ~215 lines)

• Manifest V3 Service Worker
• WebSocket client with keepalive alarm (survives MV3 Service Worker termination)
• Implements the four tools directly via Chrome APIs
• Auto-reconnects when server restarts

Accessibility Tree Builder (extension/inject/accessibility-tree.js, ~195 lines)

• Injected into target pages to build a structured DOM representation
• Outputs ref_* IDs for each interactive/structural element
• Much more token-efficient than raw HTML — AI can reason about page structure without burning context

Why accessibility tree instead of raw HTML?

A typical web page has 50-200KB of HTML. An accessibility tree of the same page is 2-10KB — 10-50x smaller. It strips away styling, scripts, and decorative elements, keeping only what matters: headings, links, buttons, inputs, and their labels.

# Raw HTML: ~150KB

  
    Settings
  

# Accessibility tree: ~50 bytes
[ref_42 link "Settings" href=/settings]

Security design

• Token auth (both layers): MCP endpoint requires Authorization: Bearer . WebSocket endpoint requires token handshake on connect. Both use the same token from ~/.browser-mcp-secrets.json.
• Minimal permissions: Only tabs, activeTab, scripting, alarms, storage. No debugger, no history, no cookies, no webRequest.
• Localhost only: Server binds to 127.0.0.1. Not exposed to the network.
• No network proxying: The extension never makes HTTP requests using your browser cookies. It only reads DOM and takes screenshots.
• On-demand: Start the server when you need it, Ctrl+C when you're done. No daemon, no auto-start.
• Auditable: ~500 lines total. Read it all in 10 minutes.

Privacy considerations

> Be aware of what you're exposing to your AI assistant.

• list_tabs sends all open tab URLs and titles to the AI assistant. If you have sensitive pages open (email, medical, private messages), their URLs and titles will be visible.
• screenshot captures whatever is on screen, including passwords, personal data, or private content.
• screenshot switches tabs: Chrome's captureVisibleTab API can only capture the active tab. If you screenshot a background tab, it will be briefly switched to the foreground.

Prompt injection risk: A malicious web page could contain hidden text that tricks your AI assistant into calling inject_script with harmful code on other tabs. Always review tool calls before approving them. Do not auto-approve inject_script calls.

Mitigations: Only run the server when you need it. Close sensitive tabs before use. Review what your AI assistant can see via list_tabs before calling other tools. Never auto-approve tool calls in your MCP client.

MV3 Service Worker survival

Chrome's Manifest V3 terminates Service Workers after ~30 seconds of inactivity. The extension uses chrome.alarms with a 24-second interval to stay alive and reconnect the WebSocket if needed.

Customization

Change the port:

MCP_PORT=9999 node index.js

Update the WebSocket URL in extension/background.js line 5 and extension/popup.html accordingly.

Add new tools:

Edit server/tools.js to register new MCP tools, and add the corresponding handler in extension/background.js's handleToolRequest switch.

How it works (for the blog post curious)

1. You start the MCP server (node index.js). It listens on :12307 for HTTP (MCP) and WebSocket (extension).
2. The Chrome Extension's Service Worker connects to ws://127.0.0.1:12307/ws.
3. Your AI assistant sends an MCP request (e.g., read_page) to http://127.0.0.1:12307/mcp with a Bearer token.
4. The server validates the token, translates the MCP call into a WebSocket message, and sends it to the extension.
5. The extension executes the corresponding Chrome API call (e.g., chrome.scripting.executeScript to inject the accessibility tree builder).
6. The result flows back: Extension → WebSocket → Server → HTTP response → AI assistant.

The whole round trip takes 100-500ms depending on page complexity.

Real-world example

See [examples/web-architectures.md](examples/web-architectures.md) for a detailed walkthrough of reading data from different web architectures — iframe-based legacy portals, React/Angular SPAs, Vue.js dynamic sites, AJAX-heavy dashboards, and CSP-strict pages.

Running tests

node --test tests/server.test.js

Tests cover: health check, token auth (missing/wrong/valid), MCP protocol handshake, and session management.

File structure

browser-mcp-lite/
├── server/
│   ├── index.js          # Fastify MCP Server + WebSocket hub
│   ├── tools.js          # Tool schemas + handlers (4 tools)
│   ├── token.js          # Shared token generation/loading
│   ├── setup.js          # One-time token generator
│   └── package.json
├── extension/
│   ├── manifest.json     # Manifest V3, minimal permissions
│   ├── background.js     # Service Worker: WS client + tool implementations
│   ├── popup.html        # Connection status UI
│   ├── popup.js
│   ├── inject/
│   │   └── accessibility-tree.js  # DOM → structured text
│   └── icons/
├── examples/
│   └── web-architectures.md    # Handling different web architectures
├── tests/
│   └── server.test.js    # Server auth + protocol tests
├── LICENSE
├── README.md             # English
└── README.zh-TW.md       # 繁體中文

Requirements

• Node.js 18+
• Chrome or Chromium-based browser
• Any MCP-compatible AI client

License

MIT

Credits

Accessibility tree approach inspired by anthropics/anthropic-quickstarts.

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: notoriouslab
• Source: notoriouslab/browser-mcp-lite
• License: MIT

Install and usage instructions live in the source repository linked above.