atlas-browser

Atlas Browser MCP Server enables visual web browsing for AI agents by labeling interactive elements on screenshots and letting you drive a browser through simple actions like navigate, click, and type. This approach helps agents interact with complex web pages while avoiding heavy DOM parsing and bot-detection hurdles.

How to use

You run the MCP server locally and connect an MCP client to control a headless or visible browser session. The client can navigate to pages, capture labeled screenshots, and perform actions by referencing element labels like [0], [1], [2], etc.

How to install

Prerequisites: you need Python 3.10+ and a compatible environment for Playwright with Chromium.

pip install atlas-browser-mcp
playwright install chromium

Configuration and usage patterns

You expose the MCP server to your client environment so the client can send commands like navigate, click, and type. Common usage patterns include setting up through a client configuration and then issuing actions step by step to guide the browser toward your target goal.

Key capabilities you can leverage in sequence include navigating to a URL, labeling interactive elements, taking labeled screenshots, clicking by label, typing into fields, and scrolling the page to reveal more content.

If you encounter a CAPTCHA, you can use the multi_click capability to select multiple labeled elements as part of the challenge.

Headless mode and viewport customization

Headless mode lets you run without a visible browser window, which is useful for servers and automation pipelines. You can disable human-like timing for faster runs if you don’t need the natural pacing.

To customize the rendered viewport, adjust the browser viewport size to your target resolution.

Troubleshooting

If Playwright reports it isn’t installed, install it and the Chromium browser again with: 
- pip install playwright
- playwright install chromium

If the browser closes unexpectedly, try running with a visible window to observe behavior: set headless to False in your client configuration.

Some dynamic pages may require extra wait time after navigation. If elements are not detected, give the page a bit more time to render before requesting actions.

Examples of typical interactions

Basic navigation and interaction follow a simple pattern: navigate to a URL, observe the labeled elements, click a label, and optionally type into a focused field. The client then requests the next labeled screenshot to confirm the new state.

Notes on integration

You can integrate with clients like Claude Desktop by configuring the MCP server in your client config. The server exposes a straightforward command interface to drive browser actions from external agents.

Security and best practices

Operate the MCP server within trusted environments, especially when handling credentials or sensitive inputs. Use proper access controls to prevent unauthorized clients from issuing commands.

Additional tips

If you encounter CAPTCHA challenges, rely on the multi_click tool to simulate human-like responses by selecting multiple labeled elements as required by the challenge.

Available tools

navigate

Load a URL and receive a labeled screenshot of the page with interactive element markers

screenshot

Capture the current page view with labels and markers

click

Click the element identified by its label ID like [0], [1], [2]...

multi_click

Click multiple elements at once to handle CAPTCHA or multi-target interactions

type

Type text into the focused input field and optionally submit the input with Enter

scroll

Scroll the page up or down to reveal more content