abrasio-mcp

MCP server that exposes Abrasio as an agentic browser for AI models.

Allows Claude, Cursor, and any MCP-compatible AI to control a real web browser with full anti-detection fingerprinting — the same stealth infrastructure used for web scraping, now driven by AI.

How it works

AI Model (Claude, Cursor, etc.)
    │  MCP tool calls
    ▼
abrasio-mcp server
    │  Abrasio SDK
    ▼
Patchright (undetected Playwright fork)
    │  CDP WebSocket
    ▼
Chrome (local stealth)  OR  Abrasio cloud worker (fingerprinted, residential IP)

The MCP server holds a single persistent browser session. The AI calls tools to navigate, observe, and interact — no configuration needed between steps.

Installation

pip install abrasio-mcp

Or install from source:

cd abrasio-mcp
pip install -e .

Modes

Local mode (free)

No API key required. Launches Chrome on your machine with Patchright stealth patches.

abrasio-mcp

Best for: development, testing, scraping sites that don't require real residential IPs.

Cloud mode (paid)

Requires an Abrasio API key. The browser runs on Abrasio cloud infrastructure with:

• Real collected browser fingerprints (not spoofed)

• Residential or datacenter IP in the target region

• Persistent profiles that accumulate browser history

ABRASIO_API_KEY=sk_live_xxx abrasio-mcp

Best for: production agents, geo-targeted tasks, heavily protected sites.

Configuration

All configuration is done via environment variables. No config files required.

Integrations

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "abrasio": {
      "command": "abrasio-mcp",
      "env": {
        "ABRASIO_API_KEY": "sk_live_xxx",
        "ABRASIO_REGION": "BR"
      }
    }
  }
}

For local mode (no API key):

{
  "mcpServers": {
    "abrasio": {
      "command": "abrasio-mcp",
      "env": {
        "ABRASIO_HEADLESS": "false"
      }
    }
  }
}

Claude Code

claude mcp add abrasio -- abrasio-mcp

With environment variables:

claude mcp add abrasio -e ABRASIO_API_KEY=sk_live_xxx -e ABRASIO_REGION=BR -- abrasio-mcp

Cursor

Add to .cursor/mcp.json in your project root:

{
  "mcpServers": {
    "abrasio": {
      "command": "abrasio-mcp",
      "env": {
        "ABRASIO_API_KEY": "sk_live_xxx"
      }
    }
  }
}

Streamable HTTP (production / remote)

Start the server:

ABRASIO_TRANSPORT=streamable-http \
ABRASIO_HOST=0.0.0.0 \
ABRASIO_PORT=8931 \
ABRASIO_API_KEY=sk_live_xxx \
abrasio-mcp

Configure the client to connect to http://your-server:8931/mcp.

Tools reference

The MCP server exposes 15 tools, organized in four groups.

Navigation

browser_navigate

Navigate to a URL and wait for the page to load.

Returns: JSON object with url (final URL after redirects), title, and status (HTTP status code).

{ "url": "https://example.com/home", "title": "Home — Example", "status": 200 }

browser_go_back

Go back to the previous page in browser history.

Returns: JSON object with url and title of the page navigated to.

browser_go_forward

Go forward to the next page in browser history.

Returns: JSON object with url and title.

browser_reload

Reload the current page.

Returns: JSON object with url and title.

browser_get_url

Get the current page URL.

Returns: String with the current URL.

Observation

browser_screenshot

Take a screenshot of the current page.

Returns: PNG image. Claude can see this image and use it to understand the visual state of the browser before deciding what to interact with.

Use this at the start of a task and after each major interaction to confirm the expected result.

browser_get_text

Extract all visible text from the current page body.

Returns: Plain text string. Useful for reading articles, prices, search results, or any text-based content without parsing HTML.

browser_get_html

Get the inner HTML of an element.

Returns: HTML string of the matched element.

browser_find_elements

Find all interactive elements on the page.

Returns: JSON array. Each item represents a clickable, fillable, or otherwise interactive element:

[
  {
    "tag": "button",
    "type": "submit",
    "text": "Sign in",
    "href": null,
    "selector": "button.login-btn",
    "x": 720,
    "y": 412,
    "visible": true
  },
  {
    "tag": "input",
    "type": "email",
    "text": "",
    "href": null,
    "selector": "input[name=\"email\"]",
    "x": 720,
    "y": 340,
    "visible": true
  }
]

Call browser_find_elements before interacting to get the correct selector for browser_click and browser_fill.

browser_wait_for

Wait for an element to appear in the DOM.

Returns: JSON object with found: true, the selector, and the element's text content.

Use this after triggering actions that cause loading states (form submissions, route changes, AJAX updates).

Interaction

All interaction tools use the Abrasio human simulation layer:

• Clicks use WindMouse — a physics-based cursor movement algorithm that mimics real hand movement with gravity, wind perturbations, and velocity management.

• Typing uses character-level timing with realistic delays, burst typing, and occasional typos followed by backspace correction.

• Scrolling uses a 3-phase easing curve (acceleration → constant → deceleration).

browser_click

Click an element by CSS selector.

Returns: Confirmation string.

Clicked: button#submit

browser_fill

Click a form input and fill it with text.

Clears any existing value, clicks the field, then types with human-like timing.

Returns: Confirmation string with character count.

Prefer browser_fill over browser_type when targeting a specific input. Use browser_type only when the input is already focused.

browser_type

Type text at the current cursor position.

Types into whichever element currently has focus. Useful for keyboard shortcuts, search boxes that open on click, or multi-step typing flows.

Returns: Confirmation string with character count.

browser_scroll

Scroll the page vertically.

Returns: Confirmation string indicating direction and distance.

browser_hover

Move the mouse over an element without clicking.

Uses WindMouse movement to reach the element. Useful for revealing dropdown menus, tooltips, or hover-triggered UI elements.

Returns: Confirmation string.

browser_press

Press a keyboard key or key combination.

Common values:

Returns: Confirmation string.

Evaluation

browser_evaluate

Execute JavaScript in the current page context.

The script can be a simple expression or a function that returns a value:

// Expression
document.title

// Arrow function
() => document.querySelectorAll('h2').length

// Function with logic
() => {
  const el = document.querySelector('#price');
  return el ? el.innerText.trim() : null;
}

// Read localStorage
() => ({ token: localStorage.getItem('auth_token') })

Returns: JSON-serialized result of the script execution.

Use this for data extraction that browser_get_text and browser_get_html cannot handle — structured data, counts, computed values, or interacting with the page's JavaScript environment.

Patterns and best practices

Starting a task

Always begin with a screenshot to understand the current state:

1. browser_navigate("https://target.com")
2. browser_screenshot()           ← see what loaded
3. browser_find_elements()        ← discover available interactions

Filling a login form

1. browser_navigate("https://site.com/login")
2. browser_screenshot()
3. browser_find_elements()        ← get selectors for email/password fields
4. browser_fill("input[name='email']", "user@example.com")
5. browser_fill("input[name='password']", "secret")
6. browser_click("button[type='submit']")
7. browser_wait_for(".dashboard")  ← wait for redirect
8. browser_screenshot()            ← confirm login succeeded

Handling dynamic content

1. browser_click(".load-more-btn")
2. browser_wait_for(".new-items-loaded")   ← wait for content
3. browser_get_text()                      ← extract updated content

Extracting structured data

1. browser_navigate("https://shop.com/product/123")
2. browser_evaluate("() => ({ name: document.querySelector('h1').innerText, price: document.querySelector('.price').innerText })")

Navigating paginated results

1. browser_navigate("https://site.com/results?page=1")
2. browser_get_text()
3. browser_click("a[aria-label='Next page']")
4. browser_wait_for(".results")
5. browser_get_text()

Architecture

abrasio-mcp/
├── pyproject.toml
└── abrasio_mcp/
    ├── __init__.py
    ├── server.py          # FastMCP server, tool registration, entry point
    ├── browser.py         # AbrasioBrowserAgent — wraps Abrasio SDK + Page
    └── tools/
        ├── __init__.py
        ├── navigate.py    # browser_navigate, go_back, go_forward, reload, get_url
        ├── observe.py     # browser_screenshot, get_text, get_html, find_elements, wait_for
        ├── interact.py    # browser_click, fill, type, scroll, hover, press
        └── evaluate.py    # browser_evaluate

Session lifecycle

The browser session is lazy-started: the MCP server process starts immediately, and the browser only opens when the first tool is called. The session persists for the lifetime of the server process — navigation history, cookies, and localStorage are preserved across all tool calls.

On SIGTERM or SIGINT, the server calls Abrasio.close() which signals the worker to stop (in cloud mode) before dropping the CDP connection, ensuring proper billing finalization.

Browser agent (browser.py)

AbrasioBrowserAgent wraps Abrasio (the SDK's unified class) and holds a single Page object. All tools delegate to this agent. The asyncio.Lock on ensure_started prevents concurrent initialization if two tools are called in rapid succession before the browser is ready.

All interaction methods use the human/actions.py primitives from the Abrasio SDK directly — no custom mouse or keyboard simulation is implemented in abrasio-mcp.

Requirements

• Python 3.10+

• mcp >= 1.0.0

• abrasio >= 0.1.2 (installs patchright automatically)

• For cloud mode: an Abrasio API key (sk_live_...)

License

MIT — Scrape Technology