abrasio-mcp
OfficialThis server exposes an agentic stealth browser over MCP, letting an AI navigate, observe, and interact with real web pages.
Navigate to URLs and manage browser history (back, forward, reload, get current URL).
Observe pages with screenshots, visible text, HTML extraction, interactive-element discovery, and waits for selectors.
Interact using human-like clicks, form filling, typing, scrolling, hovering, and keyboard presses.
Execute JavaScript in the page context to extract structured data or inspect browser state.
Maintain a persistent browser session with cookies, localStorage, and history across tool calls.
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., "@abrasio-mcp navigate to example.com and take a screenshot"
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.
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, managed stealth browser with realistic fingerprints — the same infrastructure used for web scraping, now driven by AI.
pip install abrasio-mcp # or run without installing: uvx abrasio-mcpLocal mode is free and needs no API key. For cloud mode (managed browsers with real fingerprints and persistent profiles), get a key (abr_live_...) at scrapetechnology.com/abrasio/register. Docs: scrapetechnology.com/abrasio/docs.
How it works
AI Model (Claude, Cursor, etc.)
│ MCP tool calls
▼
abrasio-mcp server
│ Abrasio SDK
▼
Patchright (Playwright fork with stealth patches)
│ 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.
Related MCP server: Scout
Installation
pip install abrasio-mcpOr 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-mcpBest 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=abr_live_xxx abrasio-mcpBest for: production agents, geo-targeted tasks, heavily protected sites.
Configuration
All configuration is done via environment variables. No config files required.
Variable | Default | Description |
| (none) | Cloud API key. If set, enables cloud mode. If unset, uses local mode. |
|
| Run browser headless. Set |
| (none) | Target region for geo-configuration, e.g. |
|
| Set |
|
| MCP transport: |
|
| Bind host for |
|
| Bind port when using |
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": "abr_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-mcpWith environment variables:
claude mcp add abrasio -e ABRASIO_API_KEY=abr_live_xxx -e ABRASIO_REGION=BR -- abrasio-mcpCursor
Add to .cursor/mcp.json in your project root:
{
"mcpServers": {
"abrasio": {
"command": "abrasio-mcp",
"env": {
"ABRASIO_API_KEY": "abr_live_xxx"
}
}
}
}Streamable HTTP (local only)
The HTTP transport has no authentication by design, so it only binds to loopback
(127.0.0.1, localhost or ::1). Anyone who could reach the port could drive the
stealth browser and spend your credits; any other ABRASIO_HOST makes the server
refuse to start.
Start the server:
ABRASIO_TRANSPORT=streamable-http \
ABRASIO_HOST=127.0.0.1 \
ABRASIO_PORT=8931 \
ABRASIO_API_KEY=abr_live_xxx \
abrasio-mcpConfigure the client to connect to http://127.0.0.1:8931/mcp.
For remote use, run the stdio transport over SSH, or keep the server on loopback and reach it through an SSH tunnel or an authenticated reverse proxy that you control.
Tools reference
The MCP server exposes 17 tools, organized in four groups.
Navigation
browser_navigate
Navigate to a URL and wait for the page to load.
Parameter | Type | Description |
|
| Full URL including scheme, e.g. |
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.
Parameter | Type | Default | Description |
|
|
| CSS selector for the target 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
}
]Field | Description |
| HTML tag name |
| Input type ( |
| Visible text, value, placeholder, or aria-label (up to 120 chars) |
| Link URL for anchor elements |
| CSS selector hint — use this with |
| Center coordinates in the viewport |
| Whether the element is within the current viewport |
Call
browser_find_elementsbefore interacting to get the correct selector forbrowser_clickandbrowser_fill.
browser_wait_for
Wait for an element to appear in the DOM.
Parameter | Type | Default | Description |
|
| (required) | CSS selector to wait for |
|
|
| Maximum wait time in milliseconds |
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.
Parameter | Type | Description |
|
| CSS selector for the element to click |
Returns: Confirmation string.
Clicked: button#submitbrowser_fill
Click a form input and fill it with text.
Parameter | Type | Description |
|
| CSS selector for the input element |
|
| Value to type into the field |
Clears any existing value, clicks the field, then types with human-like timing.
Returns: Confirmation string with character count.
Prefer
browser_filloverbrowser_typewhen targeting a specific input. Usebrowser_typeonly when the input is already focused.
browser_type
Type text at the current cursor position.
Parameter | Type | Description |
|
| Text to type |
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.
Parameter | Type | Default | Description |
|
|
| Pixels to scroll. Positive = down, negative = up. |
Returns: Confirmation string indicating direction and distance.
browser_hover
Move the mouse over an element without clicking.
Parameter | Type | Description |
|
| CSS selector for the element to hover |
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.
Parameter | Type | Description |
|
| Key name or combination |
Common values:
Key string | Action |
| Submit form / confirm |
| Move focus to next field |
| Close modal / cancel |
| Navigate dropdowns |
| Select all |
| Copy |
| Submit (macOS) |
Returns: Confirmation string.
Evaluation
browser_evaluate
Execute JavaScript in the current page context.
Parameter | Type | Description |
|
| JavaScript expression or function |
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_textandbrowser_get_htmlcannot 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 interactionsFilling 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 succeededHandling dynamic content
1. browser_click(".load-more-btn")
2. browser_wait_for(".new-items-loaded") ← wait for content
3. browser_get_text() ← extract updated contentExtracting 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_evaluateSession 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 >= 1.8, < 2(the server uses the v1FastMCPAPI)abrasio >= 0.1.2(installspatchrightautomatically)For cloud mode: an Abrasio API key (
abr_live_...) — get one at scrapetechnology.com/abrasio/register
License
MIT — Scrape Technology
Available Tools
17 toolsbrowser_clickA
Click an element using a CSS selector. Uses human-like WindMouse movement and realistic click timing. Examples: 'button#submit', '.login-btn', 'a[href="/checkout"]'
Inputs: selector (required). Returns: result
browser_evaluateA
Execute JavaScript in the current page context and return the result. The script should be a JS expression or a function body that returns a value. Examples:
'document.title'
'window.location.href'
'() => document.querySelectorAll("h2").length'
'() => { const el = document.querySelector("#price"); return el ? el.innerText : null; }' Returns the serialized result as JSON.
Inputs: script (required). Returns: result
browser_fillA
Click a form field and fill it with text using human-like typing. Clears any existing value, then types with realistic delays and occasional typos. selector: CSS selector for the input element. text: the value to type.
Inputs: text (required), selector (required). Returns: result
browser_find_elementsA
Find all interactive elements on the page (links, buttons, inputs, etc.). Returns a JSON list with each element's tag, text, type, href, selector hint, center coordinates (x, y), and whether it's visible in the viewport. Use this to discover what you can click or fill before interacting.
Inputs: no named parameters. Returns: result
browser_get_htmlA
Get the inner HTML of an element. Defaults to the full page body. Use a CSS selector to target a specific element (e.g. '#main', '.content').
Inputs: selector. Returns: result
browser_get_textA
Extract all visible text from the current page body. Useful for reading content, prices, articles, or any text-based information.
Inputs: no named parameters. Returns: result
browser_get_urlA
Return the current page URL.
Inputs: no named parameters. Returns: result
browser_go_backA
Go back to the previous page in browser history.
Inputs: no named parameters. Returns: result
browser_go_forwardA
Go forward to the next page in browser history.
Inputs: no named parameters. Returns: result
browser_hoverA
Move the mouse over an element using human-like WindMouse movement. Useful for triggering dropdown menus or tooltips before clicking. selector: CSS selector for the target element.
Inputs: selector (required). Returns: result
browser_pressA
Press a keyboard key or key combination. Examples: 'Enter', 'Tab', 'Escape', 'ArrowDown', 'Control+a', 'Meta+Return'.
Inputs: key (required). Returns: result
browser_reloadA
Reload the current page.
Inputs: no named parameters. Returns: result
browser_screenshotA
Take a screenshot of the current page. Returns a PNG image. Use this to visually understand the current state of the browser before deciding what to interact with.
Inputs: no named parameters.
browser_scrollA
Scroll the page vertically with human-like easing. pixels: positive = scroll down, negative = scroll up (default: 800).
Inputs: pixels. Returns: result
browser_typeA
Type text at the current cursor position (no element targeting). Use this after clicking an input to type into it, or to send keyboard input to the focused element. Uses human-like timing and occasional typos.
Inputs: text (required). Returns: result
browser_wait_forA
Wait for an element matching the CSS selector to appear in the DOM. Useful after triggering an action that causes a loading state. timeout: max milliseconds to wait (default 10000). Returns the element's text content when found.
Inputs: selector (required), timeout. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 17 tools
Each tool has a clearly distinct purpose: navigation, clicking, typing, scrolling, etc. No two tools overlap in function.
All tools follow a consistent 'browser_' prefix with snake_case verbs, making the naming predictable and easy to understand.
17 tools cover the full range of browser automation actions without being excessive. Each tool serves a necessary purpose.
Core browser interactions are well-covered (navigation, click, type, scroll, screenshot, etc.). Minor gaps exist for file uploads and dialog handling, but the set is largely complete.
Maintenance
Related MCP Connectors
Official Browserless MCP server: browser automation, scraping, screenshots and PDFs for AI agents.
Hosted Web Crawler MCP server for Claude, ChatGPT and Cursor, with each action switched per account.
Web scraping, browser automation, screenshots and AI data extraction for AI agents via MCP.
Unofficial Hyperbrowser MCP — wraps the Hyperbrowser AI-agent browsing API (hyperbrowser.ai).…
Related MCP Servers
- AlicenseAqualityBmaintenanceMCP server + Chrome extension that gives AI coding agents control of your real browser with existing sessions, logins, and cookies. Works with Cursor, Claude, Windsurf.24137 npm59MIT
- AlicenseBqualityDmaintenanceBrowser MCP server that connects to your existing browser, preserving sessions, passwords, and extensions, enabling AI agents to interact with web pages without bot detection.313 npm1MIT

agentify-desktopofficial
AlicenseNot gradedqualityDmaintenanceMCP server that enables AI tools to control local browser sessions for ChatGPT, Claude, and other AI services, supporting querying, navigation, file uploads, and artifact management.575Mozilla Public 2.0- AlicenseAqualityBmaintenanceMCP server for browser automation with anti-detection. Scout pages, find elements, interact with websites, and monitor network traffic from any AI client that supports the Model Context Protocol.211MIT