atlas-browser-mcp
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@atlas-browser-mcp Go to news.ycombinator.com and tell me the top 3 headlines"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
š atlas-browser-mcp
Visual web browsing for AI agents via Model Context Protocol (MCP).
⨠Features
šø Visual-First: Navigate the web through screenshots, not DOM parsing
š·ļø Set-of-Mark: Interactive elements labeled with clickable
[0],[1],[2]... markersš Humanized: Bezier curve mouse movements, natural typing rhythms
š§© CAPTCHA-Ready: Multi-click support for image selection challenges
š”ļø Anti-Detection: Built-in measures to avoid bot detection
Related MCP server: Browser MCP Server
š Quick Start
Installation
pip install atlas-browser-mcp
playwright install chromiumUse with Claude Desktop
Add to your Claude Desktop config (claude_desktop_config.json):
{
"mcpServers": {
"browser": {
"command": "atlas-browser-mcp"
}
}
}Then ask Claude:
"Navigate to https://news.ycombinator.com and tell me the top 3 stories"
š ļø Available Tools
Tool | Description |
| Go to URL, returns labeled screenshot |
| Capture current page with labels |
| Click element by label ID |
| Click multiple elements (for CAPTCHA) |
| Type text, optionally press Enter |
| Scroll page up or down |
š Usage Examples
Basic Navigation
User: Go to google.com
AI: [calls navigate(url="https://google.com")]
AI: I see the Google homepage. The search box is labeled [3].
User: Search for "MCP protocol"
AI: [calls click(label_id=3)]
AI: [calls type(text="MCP protocol", submit=true)]
AI: Here are the search results...CAPTCHA Handling
User: Select all images with traffic lights
AI: [Looking at the CAPTCHA grid]
AI: I can see traffic lights in images [2], [5], and [8].
AI: [calls multi_click(label_ids=[2, 5, 8])]š§ Configuration
Headless Mode
For servers without display:
from atlas_browser_mcp.browser import VisualBrowser
browser = VisualBrowser(
headless=True, # No visible browser window
humanize=False # Faster, less human-like
)Custom Viewport
browser = VisualBrowser()
browser.VIEWPORT = {"width": 1920, "height": 1080}šļø How It Works
Navigate: Browser loads the page
Inject SoM: JavaScript labels all interactive elements
Screenshot: Capture the labeled page
AI Sees: The screenshot shows
[0],[1],[2]... on buttons, links, inputsAI Acts: "Click
[5]" ā Browser clicks the element at that positionRepeat: New screenshot with updated labels
āāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāā
ā [0] Logo [1] Search [2] Menu ā
ā ā
ā [3] Article Title ā
ā [4] Read More ā
ā ā
ā [5] Subscribe [6] Share ā
āāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāāš¤ Integration
With Cline (VS Code)
{
"mcpServers": {
"browser": {
"command": "atlas-browser-mcp"
}
}
}Programmatic Use
from atlas_browser_mcp.browser import VisualBrowser
browser = VisualBrowser()
# Navigate
result = browser.execute("navigate", url="https://example.com")
print(f"Page title: {result.data['title']}")
print(f"Found {result.data['element_count']} interactive elements")
# Click element [0]
result = browser.execute("click", label_id=0)
# Type in focused field
result = browser.execute("type", text="Hello world", submit=True)
# Cleanup
browser.execute("close")š Requirements
Python 3.10+
Playwright with Chromium
š Troubleshooting
"Playwright not installed"
pip install playwright
playwright install chromium"Browser closed unexpectedly"
Try running with headless=False to see what's happening:
browser = VisualBrowser(headless=False)Elements not being detected
Some dynamic pages need more wait time. The browser waits 1.5s after navigation, but complex SPAs may need longer.
š License
MIT License - see LICENSE
š Credits
Built for Atlas, an autonomous AI agent.
Inspired by:
anthropic/mcp - Model Context Protocol
AskUI - Visual testing approach
Set-of-Mark prompting - Visual grounding technique
Available Tools
6 toolsclickC
Click on an element by its label ID (shown as [N] on the screenshot)
Inputs: label_id (required).
multi_clickB
Click multiple elements at once (useful for CAPTCHA, checkboxes, etc.)
Inputs: label_ids (required).
screenshotB
Take a screenshot of the current page with labeled elements
Inputs: no named parameters.
scrollC
Scroll the page up or down
Inputs: direction.
typeA
Type text at the current focus position
Inputs: text (required), submit.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelogā¦
TDQS
Scored across 6 tools
Each tool has a clearly distinct purpose with no overlap: click targets single elements, multi_click handles multiple elements, navigate loads pages, screenshot captures the current state, scroll moves the viewport, and type inputs text. The descriptions make it easy for an agent to choose the right tool for each interaction.
All tool names follow a consistent snake_case pattern with clear, imperative verbs (click, navigate, screenshot, scroll, type) and a descriptive multi_click variant. There are no deviations in style or convention, making the set predictable and readable.
With 6 tools, the server is well-scoped for browser automation, covering core interactions like navigation, clicking, typing, scrolling, and screenshot capture. Each tool earns its place without redundancy, fitting a typical range for this domain.
The toolset covers essential browser actions (navigation, interaction, viewing) effectively, with no obvious dead ends. A minor gap exists in more advanced interactions like hovering, drag-and-drop, or form submission, but agents can work around this using the provided tools for most common tasks.
Maintenance
Related MCP Connectors
AI-powered web automation. Navigate websites using AI agents for one page or a thousand
AI-powered web automation. Navigate websites using AI agents for one page or a thousand
Control a real browser from AI agents: navigate, click, type, screenshot, read the page.
Web search, browser automation, scraping, crawling and CAPTCHA solving for AI agents.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceEnables AI agents to interact with web browsers using natural language, featuring automated browsing, form filling, vision-based element detection, and structured JSON responses for systematic browser control.64MIT
- AlicenseNot gradedqualityDmaintenanceEnables AI agents to control browsers with human-like behavior, stealth anti-detection, and 70 tools for navigation, interaction, and monitoring.8 npm1MIT
- FlicenseNot gradedqualityBmaintenanceEnables AI agents to autonomously control browsers with zero-hardcoding semantic DOM interaction, real-time network telemetry, Cloudflare/bot self-healing, and human-in-the-loop reasoning for web QA and automation.-
- AlicenseBqualityCmaintenanceEnables LLM agents to automate web browsers with token-efficient perception, stealth anti-detection, captcha solving, network interception, and isolated persistent sessions.51MIT