@nxavis/agent-browser-mcp
# @nxavis/agent-browser-mcp
[](https://www.npmjs.com/package/@nxavis/agent-browser-mcp)
[](https://opensource.org/licenses/MIT)
[](https://nodejs.org/)
A [Model Context Protocol (MCP)](https://modelcontextprotocol.io/) server that provides browser automation capabilities through [Vercel's agent-browser](https://github.com/vercel-labs/agent-browser). This enables LLMs to interact with web pages using a fast Rust CLI with Node.js fallback.
## Quick Start
```bash
# 1. Install agent-browser CLI
npm install -g agent-browser && agent-browser install
# 2. Add to Claude Desktop (or your MCP client)
npx @nxavis/agent-browser-mcp
```
Then use tools like `browser_navigate`, `browser_click`, `browser_snapshot` to control the browser from your AI agent.
## Features
- **AI-Optimized Browser Control** - Semantic element locators using accessibility properties, text matching, and data attributes
- **Session Isolation** - Multiple isolated browser sessions with separate cookies, storage, and navigation history
- **Comprehensive Automation** - Navigation, form filling, clicking, scrolling, keyboard input, and more
- **Data Extraction** - Get text, HTML, attributes, accessibility snapshots, screenshots, and PDFs
- **Cookie Management** - Full control over browser cookies and storage
- **JavaScript Execution** - Run arbitrary scripts in the browser context
- **Network Inspection** - Monitor console messages and network requests
## Installation
```bash
npm install @nxavis/agent-browser-mcp
```
Or run directly with npx:
```bash
npx @nxavis/agent-browser-mcp
```
### Prerequisites
- Node.js 18 or newer
- [agent-browser](https://github.com/vercel-labs/agent-browser) CLI installed:
```bash
# Install agent-browser globally
npm install -g agent-browser
# Download Chromium browser
agent-browser install
# On Linux, install system dependencies if needed:
# agent-browser install --with-deps
```
> **⚠️ Windows Note:** agent-browser currently has [known issues on Windows](https://github.com/vercel-labs/agent-browser/issues/56) with native shells (PowerShell/CMD). For Windows users, we recommend using [WSL (Windows Subsystem for Linux)](https://learn.microsoft.com/en-us/windows/wsl/install) until the upstream issue is resolved.
## Configuration
### Claude Desktop
Add to your `claude_desktop_config.json`:
**macOS:** `~/Library/Application Support/Claude/claude_desktop_config.json`
**Windows:** `%APPDATA%\Claude\claude_desktop_config.json`
```json
{
"mcpServers": {
"agent-browser": {
"command": "npx",
"args": ["@nxavis/agent-browser-mcp"]
}
}
}
```
### VS Code
Add to your VS Code settings (JSON):
```json
{
"mcp": {
"servers": {
"agent-browser": {
"command": "npx",
"args": ["@nxavis/agent-browser-mcp"]
}
}
}
}
```
### Antigravity
Add to your `.mcp.json` file in your project root or home directory:
```json
{
"mcpServers": {
"agent-browser": {
"command": "npx",
"args": ["@nxavis/agent-browser-mcp"]
}
}
}
```
Or use the global configuration at `~/.mcp.json`:
```json
{
"mcpServers": {
"agent-browser": {
"command": "npx",
"args": ["@nxavis/agent-browser-mcp"],
"env": {
"AGENT_BROWSER_PATH": "/usr/local/bin/agent-browser"
}
}
}
}
```
### Custom agent-browser Path
If `agent-browser` is not in your PATH, specify its location:
```json
{
"mcpServers": {
"agent-browser": {
"command": "npx",
"args": ["@nxavis/agent-browser-mcp"],
"env": {
"AGENT_BROWSER_PATH": "/path/to/agent-browser"
}
}
}
}
```
## Available Tools
### Navigation
- `browser_navigate` - Navigate to a URL
- `browser_go_back` - Navigate back in browser history
- `browser_go_forward` - Navigate forward in browser history
- `browser_reload` - Reload the current page
### Interaction
- `browser_click` - Click on an element
- `browser_fill` - Fill a text input field
- `browser_type` - Type text character by character
- `browser_hover` - Hover over an element
- `browser_scroll` - Scroll the page or a specific element
- `browser_select` - Select an option from a dropdown
- `browser_check` - Check a checkbox or radio button
- `browser_uncheck` - Uncheck a checkbox
- `browser_press` - Press a keyboard key
### Data Extraction
- `browser_get_text` - Get text content from an element or page
- `browser_get_html` - Get HTML content
- `browser_get_attribute` - Get an attribute value
- `browser_get_url` - Get the current page URL
- `browser_get_title` - Get the current page title
- `browser_snapshot` - Get accessibility tree snapshot
### Session Management
- `browser_new_session` - Create a new isolated browser session
- `browser_close_session` - Close a browser session
### Screenshots & PDF
- `browser_screenshot` - Take a screenshot
- `browser_pdf` - Generate a PDF of the current page
For a complete list of tools and their parameters, see the [original documentation](https://github.com/vercel-labs/agent-browser).
## Development
```bash
# Clone the repository
git clone https://github.com/nxavis/agent-browser-mcp.git
cd agent-browser-mcp
# Install dependencies
npm install
# Build
npm run build
# Run tests
npm test
# Watch mode
npm run dev
```
## License
MIT
## Credits
This package is based on [agent-browser-mcp](https://github.com/minhlucvan/agent-browser-mcp) by minhlucvan, adapted for the @nxavis ecosystem.
TDQS
Scored across 34 tools
Each tool targets a distinct browser action or query, with clear separation between navigation, interaction, extraction, session, and diagnostic tools. Slight overlaps like fill vs type are clearly differentiated by behavior, and get_text vs get_html vs snapshot have distinct outputs.
All tools uniformly use the browser_ prefix and snake_case with action-oriented verbs. The mix of get_* and direct verbs is consistent with the domain, and patterns like wait_for_* and is_* follow predictable conventions.
At 34 tools, the set exceeds the 25+ threshold for 'too many'. While each tool serves a purpose, several could be consolidated (e.g., is_visible/enabled/checked into a single state query, or get_text/html/attribute into a content extractor), making the surface heavier than necessary.
The tools cover the full browser automation lifecycle: navigation, interaction, content extraction, waiting, session management, cookies, JavaScript execution, and diagnostics (console/network). Minor gaps like file upload/download are non-core and do not create dead ends.