What if I still get flagged or banned?

No browser can promise otherwise. Detection also depends on your proxy quality, your behaviour and the account history you bring, and any vendor claiming a guarantee is selling you something. What we do promise: the fingerprint is coherent and checkable - run the public detectors yourself before you trust a profile with anything valuable, and start on the free tier so that check costs you nothing.

The antidetect browser your AI agent can drive

Q: Do I have to rewrite my Playwright scripts?

No. launch() returns standard Playwright Page and BrowserContext instances. Existing scripts work out of the box - only the launch line changes.

Q: How does MCP server mode work?

Run npx anti-detect-browser --mcp and the SDK exposes launch, navigate, click, type, read and screenshot tools. Drop it into Claude or any MCP client and your agent operates a fully fingerprinted browser.

Real fingerprints.
Agent-ready control.

Identities that survive re-login, data that never leaves your disk unless you say so, and detector results you can reproduce yourself.

Not spoofed. Real devices.

Every profile carries the fingerprint of a real device - not randomized noise that anti-bot systems flag. Consistent across Canvas, WebGL, fonts and the engine layer.

CanvasWebGLFontsEngine-level

Chrome 153 is already here

A profile that reports a Chrome version nobody is running any more is a profile that stands out. We rebuild on each new Chromium major and publish it within days, so your identities are as current as the browsers around them. Older majors stay downloadable and selectable per profile - an aged account keeps the version it was aged on.

Chrome 153Days, not quartersOlder majors keptPer-profile choice

Playwright you already wrote - or one command

Standard Playwright API, no proprietary SDK to learn: connect your agent or script in one line and it just runs. Point an MCP client at it instead and the agent drives a profile directly. And when you only want the data, a recipe is one command - recipe run reddit/hot returns JSON, and fanout runs that same command across N isolated identities at once. Recipes →

PlaywrightMCPRecipesfanout

Passkeys, captured and replayed

Every profile carries its own virtual authenticator. When a site enrols a passkey, AntiBrow captures it automatically and replays it on the next sign-in - passwordless logins stay warm per identity, no re-enrolment. On by default for every new profile. How passkeys work here →

WebAuthnFIDO2Auto-captureOn by default

Residential proxies already wired in

Every paid plan includes managed residential proxies - bind one to a profile and timezone, locale and geo follow the exit IP. Bring your own HTTP or SOCKS5 instead if you'd rather; credentials are answered inside the engine, so no helper extension is loaded and chrome://extensions stays empty. Managed proxies →

Included on paid plansOr bring your ownAuth in the engineNo proxy extension

A relay with no proxy protocol to recognize

SOCKS5 and HTTP CONNECT announce themselves before your first byte moves. The AntiBrow relay doesn't: the engine speaks it natively over one WebSocket, sealing every frame with AES-256-GCM, so there is no CONNECT line, no SOCKS5 handshake and no fixed-length header to match on. It's MIT-licensed and meant to be self-hosted on a domain you control. How the relay works →

No proxy signatureAES-256-GCM per frameOpen sourceSelf-hostable

Android device profiles, no phone required

Set deviceType: 'android' on a profile and it simulates a full Android phone - user agent, Client Hints, touch, screen - entirely on the Windows, macOS or Linux machine you already have. Nothing installs on an actual device, because none is involved. Free on every plan, including Free.

Simulated, on your desktopNo phone neededFree, every plan

Log in once. Stay logged in - on your own disk.

Cookies, storage and passkeys persist across restarts, so an account you warm up once stays warm. All of it is written to your machine, and cloud sync is opt-in per profile - a profile you don't sync is a profile we can't read. Create as many identities as you want at zero cost: paid plans buy concurrency, sync and proxies, never the right to make one more profile.

Cookies & sessionsLocal-firstOpt-in syncPortable export$0 forever

Watch it live, then hand it over

Stream a running session to your dashboard in real time - no black box about what your agents and accounts are doing. When someone else needs that environment, there are three ways to pass it on: export it as a file, share a link to it, or transfer ownership outright. Only one of the three can be revoked afterwards. Which one to use →

Live ViewExport a fileShare a linkTransfer ownership

Next: Everything API. An account signed in inside a synced profile, reachable as an HTTP call - same session, same fingerprint, same proxy exit you use by hand. In internal testing right now. Get notified at launch →

Deep dive · Real-device fingerprints

Every signal lines up - because it came from one real machine.

Randomized fingerprints betray themselves: a Windows user-agent paired with a font list no matching machine would have, or a GPU string no real card ever shipped. Anti-bot systems are built to catch exactly those contradictions.

We sample each profile whole, from a coherent device profile, and spoof it at the kernel level through the Engine runtime. Canvas, WebGL, audio, fonts and the protocol layer all agree - by construction, not by luck.

30+ categories, 500+ parameters - Canvas, WebGL, WebGPU, audio, fonts, WebRTC, battery, media devices and more.

Coherent by construction - every value drawn from the same real device, never mixed and matched.

Kernel-level spoofing - a battle-tested engine for Windows, macOS & Linux that spoofs below the JS layer, not just via injected JS.

fingerprint · profile "ads-us-east" real device

User-Agent Win64 · Chrome 153.0

Platform Win32

WebGL renderer ANGLE · Intel Iris Xe

Canvas hash a1f3·9e7c·2b

Audio context 124.0438·52

Fonts 218 Windows fonts

Timezone / IP America/New_York

Hardware 8 cores · 16 GB

All 30+ categories consistent0 contradictions

Proof · Real detector & anti-bot tests

Open the detectors.
Walk through the bot walls.

Verified by launching a real profile through a residential proxy, then opening the checks anyone can run - the public fingerprint detectors Whoer.net and CreepJS, plus live enterprise anti-bot walls from DataDome and Kasada. Live capture, not mockups.

Whoer.net result - 100% disguise, proxy not detected, no DNS leak — **Whoer.net** - proxy not detected, no DNS leak, and OS, browser & geo all line up. This shot predates the current kernel; the 2026-08-02 run repeats it on Chrome 150.

CreepJS result - 0% headless, 0% stealth, high confidence — **CreepJS** - high confidence, `webdriver=false`, timezone & locale aligned to the proxy.

leboncoin.fr, protected by DataDome, served its real listings with no CAPTCHA wall — **DataDome** - leboncoin.fr (`x-datadome: protected`) served its real listings; no CAPTCHA wall, no `captcha-delivery` redirect.

footlocker.com, protected by Kasada, after the proof-of-work challenge was solved in-browser and the real storefront loaded — **Kasada** - footlocker.com (`x-kpsdk-*`) threw its 429 proof-of-work challenge; the profile solved it in-browser and the real storefront rendered.

How this run was made - so you can repeat it

Create a fresh profile in the desktop app, accept the fingerprint it generates, change nothing else.
Bind a managed residential proxy; timezone, locale and geo follow the exit IP automatically.
Open each target in that profile and screenshot the result unedited - the four shots above.

Every check here is public: whoer.net and abrahamjuliot.github.io/creepjs report on themselves, and the DataDome and Kasada walls are the live production defences on leboncoin.fr and footlocker.com. Nothing here needs our cooperation to verify, and the free tier is enough to run all four yourself.

Detection is an arms race and results move - exact scores vary with proxy, target and fingerprint. We publish runs, not guarantees: every dated run is here, including the ones we fail.

playwright.ts playwright.py mcp.json

// Launch a real-device profile, get Playwright back
import { AntiDetectBrowser } from "anti-detect-browser"

const ab = new AntiDetectBrowser({ key: "your-api-key" })

const { browser, page } = await ab.launch({
  profile: "ads-us-east",
  fingerprint: { tags: ["Windows 10", "Chrome"] },
  proxyId: "us-east-01",
})

await page.goto("https://example.com")
// → standard Playwright Browser + Page
# Same product, same profiles - from Python
from antibrow import launch

browser = launch(
    profile="ads-us-east",
    proxy="http://user:pass@gate.example.com:8080",
)

page = browser.new_page()
page.goto("https://example.com")
# → standard Playwright Page, plus browser.cdp_url
// Drop into any MCP client - Claude included
{
  "mcpServers": {
    "anti-detect-browser": {
      "command": "npx",
      "args": ["anti-detect-browser", "--mcp"],
      "env": {
        "ANTI_DETECT_BROWSER_KEY": "your-api-key"
      }
    }
  }
}
// → agent gets launch, navigate, click, screenshot tools

Point your agent and go.

Hand any AI agent a fully fingerprinted browser over MCP, or wire it up with the standard Playwright API you already write - in TypeScript or Python, with no proprietary SDK to learn.

One-line MCP setup. Run the SDK as an MCP server with browse, click, type, read and screenshot tools.

Standard Playwright handles. No bespoke API to learn - browser and page behave exactly as you expect.

Headful or headless. Same fingerprint integrity whether a human watches via Live View or an agent runs solo.

Already patching Playwright yourself? Stealth plugins are free and they work - right up until the release that breaks them, and then the arms race is your on-call rotation. AntiBrow spoofs in the engine instead of patching from JavaScript, ships profile storage, proxy binding and passkey persistence you'd otherwise build, and the free tier costs the same as your current stack. See how it compares →

Questions, answered.

Each fingerprint is sampled as a coherent device profile - actual WebGL renderers, canvas hashes, audio stacks and font lists. Nothing is randomly generated, so signals never contradict each other.

Not unless you turn cloud sync on for that profile. Profiles are created on your own disk and stay there: cookies, localStorage, IndexedDB and passkeys are all local files. Cloud sync is opt-in per profile and available on paid plans; a profile with sync off never leaves the machine. Deleting a synced profile removes the cloud copy too. See exactly what crosses the network →

No. launch() returns standard Playwright Page and BrowserContext instances. Existing scripts work out of the box - only the launch line changes.

Run npx anti-detect-browser --mcp and the SDK exposes launch, navigate, click, type, read and screenshot tools. Drop it into Claude or any MCP client and your agent operates a fully fingerprinted browser.

Local profiles keep working. They are ordinary directories on your disk, launched by a browser binary already installed on your machine, and the free tier has no expiry - so downgrading only lowers your concurrency cap. Every profile also exports to a portable archive file you keep.

Partly, today. AntiBrow imports and exports its own portable profile archives, which carry the fingerprint, cookies, storage, bookmarks and passkeys, so moving between machines is lossless. Direct importers for other vendors’ formats are not shipped yet - tell us which one you use and we’ll prioritise it.

Yes. Cookies, localStorage, IndexedDB and session state are stored per profile, so logged-in accounts stay warm across restarts.

Yes. Every profile gets its own virtual authenticator, so passkeys enrolled on a site are captured automatically and replayed on later logins - passwordless sign-ins persist per identity across restarts, just like cookies. Passkey capture & replay is on by default when you create a profile, and can be turned off per profile.

Yes, SOCKS5 included. The credentials are answered inside the engine: HTTP and HTTPS 407 challenges in the network stack, SOCKS5 by RFC 1929 username/password negotiation. Nothing is installed to make that work, so chrome://extensions stays empty - a proxy-auth helper extension sitting there is itself a tell, and it is one most antidetect browsers still leave behind.

No browser can promise otherwise. Detection also depends on your proxy quality, your behaviour and the account history you bring - any vendor claiming a guarantee is selling you something. What we do promise is that the fingerprint is coherent and checkable: run the public detectors yourself before you trust a profile with anything valuable, and start on the free tier so that check costs you nothing.

Windows, macOS and Linux are all supported. The desktop app, the MCP server and both automation SDKs run on all three - Node 18+ for the TypeScript SDK, Python 3.9+ for the Python one.

Yes - unlimited local profiles with the full fingerprint engine, no credit card required. Paid plans start at $9.9/mo and add concurrency, cloud sync, managed proxies and Live View.

The antidetect browser
your AI agent can drive.