-
Notifications
You must be signed in to change notification settings - Fork 221
ghost cursor human mouse
Two things get a session flagged, and they are not the same thing. One is what the browser reports about itself: the GPU, the fonts, the screen, the driver flags. The other is what the browser does once it is on the page: where the pointer goes, how fast, and whether it ever moves at all before a click.
ghost-cursor addresses the second one. It generates human-like mouse paths using Bezier curves with realistic acceleration and overshoot, so a click arrives at a button along a plausible arc instead of appearing on it. This page is about how that layer fits with a browser that is already real at the fingerprint level, and, just as important, what it does not fix.
invisible_playwright makes the browser look like a real Firefox driven by a real person. The patch lives in the C++ engine, so the values a detector reads come out consistent with each other: the GPU string matches the pixels it draws, the fonts match the platform, the driver flags read like a browser nobody automated. That is why it passes most fingerprint, TLS and driver checks. A page-level property patch cannot claim the same, because it is editing answers after the engine has already given a different one.
None of that is a statement about behaviour. Fingerprint realness answers "what is this browser"; behaviour realness answers "is a person driving it". They are measured by different code, they fail independently, and you need both. A perfectly consistent fingerprint that teleports its pointer from corner to corner is a real browser being driven by something that is obviously not a person, and some sites watch only for that.
Switching from stock Playwright is two lines, and the pointer behaviour changes with it:
from invisible_playwright import InvisiblePlaywright
with InvisiblePlaywright(seed=42) as browser:
page = browser.new_page()
page.goto("https://example.com")
page.click("#submit") # the pointer arcs to the button on a Bezier curveThe browser object is a real Playwright Browser, so every method you already use
works unchanged. What is different is that a page.click() routed through this class
does not jump the cursor onto the target: a built-in cursor engine moves it there
along a curved path first. You can select which engine drives that motion, or turn it
off, with an environment variable:
# python (default) drives the built-in path engine; "off" hands motion back to you
export INVPW_CURSOR_ENGINE=offSo for the ordinary case, click paths are already handled and you do not have to add anything. The reason this page exists is that click paths are one slice of pointer behaviour, not all of it.
The built-in arc covers the pointer motion for clicks the class drives. Several common patterns fall outside that:
-
Moving without clicking. If you call
page.mouse.move(x, y)directly, that is a raw Playwright call and it teleports: Playwright's own docs say the default single-step call "emits a singlemousemoveevent at the destination location," which is exactly the jump a detector can catch. Exploratory drift, moving toward an element to trigger a hover state, or repositioning between actions are all straight jumps unless something draws them. - Hover before click. A person's pointer is usually near a control before they commit to it. A driver that only ever moves at the moment of the click has a different signature from one that approaches first.
- The shape and timing you want. If you need a specific approach, an overshoot and correction, or a particular pace, you want to own the path rather than accept a default.
For all of these you set INVPW_CURSOR_ENGINE=off so the two engines are not both
trying to move the mouse, and you drive the pointer yourself. ghost-cursor is a
convenient way to do that, because it produces the curved, accelerating motion for
arbitrary moves and not only for clicks.
ghost-cursor started as a Node library; the Python port python-ghost-cursor drives
a Playwright page directly. Because invisible_playwright hands you a genuine
Playwright Page, the port attaches to it with no adapter:
pip install invisible-playwright python-ghost-cursorimport os
from invisible_playwright import InvisiblePlaywright
from python_ghost_cursor.playwright_sync import create_cursor
# hand pointer motion to ghost-cursor so the two engines do not fight over the mouse
os.environ["INVPW_CURSOR_ENGINE"] = "off"
with InvisiblePlaywright(seed=42) as browser:
page = browser.new_page()
page.goto("https://example.com")
cursor = create_cursor(page)
cursor.move("#menu") # curved approach, no click, sets up a hover
cursor.click("#submit") # Bezier path to the target, then the clickThe identity stays reproducible: seed=42 gives the same GPU, canvas hash, audio
context and fonts on every run, so if a session is flagged you can replay it and tell
a change in the site from a change in the machine. The pointer path from ghost-cursor
is not seed-locked to that identity, but the fingerprint underneath it is, which is
the part that has to stay stable for a bisect to mean anything.
If you would rather keep the built-in engine for clicks and add ghost-cursor only for
the moves between them, leave INVPW_CURSOR_ENGINE at its default and use the cursor
for move calls alone. Either arrangement is fine; what you must avoid is two engines
animating the same click, which produces a doubled or contradictory motion that is
worse than either alone.
A human-shaped mouse path removes one tell. It does not remove the others, and it is worth being explicit so a smooth cursor is not mistaken for a solved session.
- IP reputation. A real browser with a lifelike pointer, arriving from a datacenter range or a public proxy everyone else is using this minute, still loses. The pointer path is invisible to the part of the system that scores the address. You supply a clean exit; see configuration for how the proxy is set and how the timezone is derived from it.
- Pacing and rate. ghost-cursor shapes a single motion. It says nothing about how many actions you take, how close together, or whether a form is filled in eighty milliseconds. Per-account quotas and rate limits are counted server-side and no cursor library touches them. Human pacing between actions is yours to add.
- Session rhythm overall. Scrolling, reading pauses, the gap between landing and the first interaction, and for automated agents the pause shaped like model latency are all behaviour that a mouse-path library does not cover.
The rule from the rest of this site holds here: stealth means looking real, not suppressing a signal. A pointer that never moves is not a neutral absence, it is a signal, in the same way that a trust score is built from many surfaces agreeing rather than from any one of them being silent. ghost-cursor fills one of those surfaces. It does not fill the rest, and no honest page would claim it makes a session undetectable.
Fingerprint realness and behaviour realness are separate problems with separate fixes. invisible_playwright covers the first at the engine level and already arcs the pointer for the clicks it drives; ghost-cursor lets you own the full pointer path when you need motion beyond those clicks. Layer them by handing motion to one engine at a time, keep the seed fixed so failures are reproducible, and remember that a lifelike cursor is one honest improvement among several the reader still has to supply: a clean exit, human pacing, and a session rhythm that is not all one speed.
Does invisible_playwright already move the mouse on a curve? Yes. A page.click()
routed through the class arcs the pointer to the target on a Bezier path by default.
ghost-cursor is for pointer motion beyond those clicks, or when you want to own the
path yourself.
Do I still need ghost-cursor then? Only for what the built-in click arc does not cover: moving without clicking, hovering before a click, or a specific shape and pace. For plain clicks you do not have to add anything.
How do I stop the two cursor engines fighting? Set INVPW_CURSOR_ENGINE=off
before launch when ghost-cursor is driving the pointer, so only one engine animates
the mouse. Never let both animate the same click.
Does a human mouse path make me undetectable? No. It removes the teleport-click tell and nothing else. IP reputation, rate limits, per-account quotas and overall timing are separate, and you still supply a clean proxy and human pacing.
Does ghost-cursor need a special browser build? No. invisible_playwright returns a
standard Playwright Page, so python-ghost-cursor attaches to it the same way it
would to any Playwright page.
Will the pointer path be the same for a fixed seed? The fingerprint is seed-locked and reproduces exactly; the ghost-cursor path is not tied to the seed. The identity underneath stays stable, which is what a bisect depends on.
- This project's quickstart and configuration pages for the real launch API, the
built-in cursor engine, and the
INVPW_CURSOR_ENGINEvariable. -
Playwright's own mouse API documentation
for how
page.mouse.move()behaves with no engine layered on top of it. - The public ghost-cursor project and its Python port for the Bezier-path pointer motion, read from their own documentation.
- The release notes for this site's separation of fingerprint surfaces from behavioural ones, which is why pointer realism is documented as a layer you add rather than a claim the engine makes for you.
See also: what a browser trust score actually measures, stealth for AI browser agents for the timing tells a cursor path does not cover, and the automation layer for why the driver itself is a surface.
Written while maintaining invisible_playwright, a Firefox patched at the C++ level driven by stock Playwright. It looks like a real browser and already arcs the pointer for its clicks; the rest of behaviour, and a clean exit, are still yours to supply.
Documentation
Guides
-
Browser Identity
- navigator.webdriver is not the tell you think it is
- hardwareConcurrency, deviceMemory and storage quota
- Screen size and viewport tells in headless browsers
- Playwright headless vs headed: what detectors see
- Playwright User Agent: Why You Should Not Set It
- Client Hints and Sec-Fetch: headers that must agree
- Codec fingerprinting: canPlayType and MediaCapabilities
- Permissions API: the two answers that must agree
- CSS fingerprinting: what media queries reveal
- What privacy.resistFingerprinting actually does
- speechSynthesis.getVoices() returns an empty array
- Browser extensions are a fingerprint surface
- BFCache and pageshow.persisted under browser automation
- Service workers, storage partitioning and automation
- Web Workers: where page-level fingerprint patches fail
- fake-useragent is archived: what changes and what doesn't
- navigator.buildID and the stale build date tell
- navigator.maxTouchPoints and pointer consistency
- navigator.platform and oscpu on a spoofed OS
- navigator.vendor and productSub: the Firefox tells
- Accept-Language header vs navigator.languages
- window.devicePixelRatio: the pref that spoofs it
- Can you be fingerprinted in incognito mode?
- Is changing the user agent enough to avoid detection?
- Can a website tell you are running on a server?
- Can two devices share a browser fingerprint?
- Does clearing cookies stop fingerprint tracking?
- Color-gamut and HDR media queries as a fingerprint
- Battery API fingerprint: does Firefox expose it?
- Is navigator.connection a fingerprint in Firefox?
- Can the Gamepad API fingerprint or detect a bot?
- Do accelerometer and gyroscope APIs leak on desktop?
- prefers-reduced-motion and other OS-setting tells
- Does storage quota estimate reveal disk size?
- Can scrollbar width reveal my operating system?
-
Canvas, WebGL, Fonts and Audio
- Canvas fingerprint noise: why per-call randomising fails
- Firefox WebGL renderer strings: what ANGLE reports
- WebGL parameters: the numbers are the same on every GPU
- Your renderer string says NVIDIA. Your pixels say software.
- Why headless browsers render different fonts
- How to make Linux and macOS report real Windows fonts
- measureText and TextMetrics as a fingerprinting surface
- AudioContext fingerprinting, and why adding noise backfired
- Canvas and WebGL fingerprints, identical across OSes
- Emoji fingerprinting: why emoji look the same on any OS
- Detecting installed fonts in JavaScript by width
- WebGL shader precision as a fingerprint surface
- AudioContext sampleRate and latency as a fingerprint
- Is WebGPU a browser fingerprint?
-
Network, Proxy and WebRTC
- WebRTC leak with a proxy in Playwright and Selenium
- WebRTC ICE candidate spoofing: the fields that give it away
- Playwright proxy in Python: per-context, and what leaks
- Playwright proxy not working? SOCKS5 auth in Python
- Playwright timezone does not match the proxy IP
- JA3 and JA4: why a TLS fingerprint cannot be patched
- Playwright in Docker: it runs, and still gets blocked
- Web scraping keeps getting blocked with good proxies
- Python web scraping blocked? The TLS fingerprint reason
- SOCKS5 vs HTTP proxy: what each does in the browser
- WebRTC IPv6 leak: why a proxy does not stop it
- HTTP/2 fingerprint: the layer above the TLS handshake
- TLS fingerprint vs User-Agent: the contradiction
- WebRTC has no ICE candidates behind a proxy
- WebRTC IP that matches the proxy exit, by design
- How to check if a proxy leaks your real IP
- about:webrtc: read your real ICE candidates
- Offline timezone resolution from a proxy exit IP
- Residential vs datacenter vs mobile proxies explained
- Sticky vs rotating proxy sessions: which to use
- Does a proxy leak DNS? DoH and DNS leaks explained
- HTTP/3 and QUIC fingerprint: what a site sees
- What is ASN and IP reputation in bot detection?
- What does a mobile carrier IP look like to a site?
- IPv6 vs IPv4: which does your proxy expose?
- Geolocation API vs IP location: keep them consistent
- Does chaining two proxies help avoid detection?
-
The Automation Layer
- Function.prototype.toString and the [native code] check
- The ChromeDriver
cdc_variable, and why renaming it fails - Why an attached debugger makes automation detectable
- Execution context was destroyed, and when it means detection
- Human-like mouse movement: Bezier curves are the easy part
- Why a Playwright upgrade broke 97 of 133 tests overnight
- Playwright persistent profile: what it fixes and breaks
- Why humanized mouse movement can fail on hover()
- Why content_frame() returns None for a cross-origin iframe
- Orphaned Firefox processes on Windows: the killed-runner leak
- Firefox launches but Playwright can't drive it: packaging gap
- Why automating login is riskier than reusing a session
- Playwright new_page vs new_context: the viewport tell
- Playwright dialog and popup handling without a tell
- Playwright download files with Firefox and the tell
- Playwright connect_over_cdp does not work with Firefox
- Playwright mobile emulation on Firefox and isMobile
- Playwright isTrusted: are automated clicks real?
- Playwright set_input_files uploads and the tell
- Can websites detect Playwright? What is actually visible
- Does Playwright Set navigator.webdriver to True?
- Does Playwright Leave Traces a Website Can See?
- Does Playwright Change My Browser Fingerprint?
- Can I Use My Real Browser Profile With Playwright?
- Does Playwright Support Firefox Stealth?
- Is Playwright Firefox Harder to Detect Than Chromium?
- Does Playwright Get Detected on the First Request?
- Why Playwright's bundled Firefox is easy to detect
- ghost-cursor human mouse paths with Playwright
- Stock Playwright, patched Firefox: how they connect
- Intercept and mock network requests with page.route
- Record and replay HTTP traffic with HAR in Playwright
- Record a Playwright trace to debug a failed scrape
- Record a video of a Playwright browser session
- Save and reuse login with storage_state in Playwright
- Read and set cookies in a Playwright context
- Set geolocation and permissions per Playwright context
- Handle HTTP basic auth in Playwright (http_credentials)
- Isolate identities with a browser context per session
- Drag and drop elements in Playwright with drag_to
- When to use an HTTP client vs a real browser
- Migrating from requests + BeautifulSoup to a browser
-
AI Agents and Frameworks
- AI browser agents and stealth: what fits and what does not
- browser-use gets detected: what you can and cannot change
- crawl4ai stealth mode and custom browser engines
- Give a LangChain agent an invisible_playwright browser
- Feed invisible_playwright pages into a RAG index
- Computer-use agents and browser fingerprint detection
- Give an MCP browser server a stealth Firefox engine
- Give each AI agent a reproducible browser identity
- Run parallel browser agents with distinct fingerprints
- Why AI browser agents have their own timing signal
- Running an AI browser agent headless on a server
- Give a browser agent a persistent logged-in session
- smolagents: hand the agent an invisible_playwright tool
- Stagehand and stealth: why a Firefox engine won't drop in
- DOM-reading vs screenshot agents: which stealth helps
- Back a computer-use agent with a real browser engine
- AI agent retry loops trip rate limits, not fingerprints
-
Detectors, Explained
- What bot.sannysoft.com actually checks, row by row
- How CreepJS decides you are lying
- What BotD actually detects, and what it does not
- Why a FingerprintJS visitor ID changes
- reCAPTCHA v3 score: why a fresh browser scores badly
- BrowserLeaks canvas and WebGL hash, explained
- What BrowserLeaks actually tests, surface by surface
- Browser trust scores explained: what the number means
- How do websites detect bots?
- What is a browser fingerprint?
- What data does a website collect about your browser?
- Does a VPN stop browser fingerprinting?
- Do websites know you are using a script?
- How accurate is browser fingerprinting?
- Can a website detect a virtual machine?
- Can websites detect a datacenter or proxy IP?
- getClientRects fingerprinting: subpixel geometry as ID
- Notification.permission as a bot-detection signal
- speechSynthesis voices as a cross-platform fingerprint
- Can a website detect typing by keystroke timing?
- Can a website detect Clipboard API access?
- What are mouse-dynamics behavioural biometrics?
-
Testing and Troubleshooting
- How to test bot detection without a false pass
- Playwright detected as a bot: the checklist to fix it
- Firefox preferences that silently do nothing
- Slow browser launch: a per-request timeout is not a budget
- Playwright screenshot returns noise: readback fix
- Canvas fingerprint changes every run: use a seed
- Playwright TargetClosedError: the causes and the fixes
- Why am I blocked with a clean fingerprint?
- Why Does My Playwright Script Get Blocked?
- Is Playwright headless detectable? What sites check
- Can You Run Playwright Without Being Detected?
- Why Playwright Works Locally but Fails in the Cloud
- Does Playwright Trigger reCAPTCHA More Often?
-
Scraping with Playwright
- How to scrape without getting blocked
- How to scrape a site that blocks headless browsers
- How to scrape infinite scroll pages with Playwright
- How to rotate proxies when scraping with Playwright
- How to scrape data behind a login with Playwright
- How to run Playwright in Docker without getting detected
- How to use invisible_playwright in Docker
- Playwright bot detection: how to avoid it in Python
- How to scrape paginated pages with Playwright
- How to download files with Playwright
- How to upload files with Playwright, and verify it landed
- How to handle cookie consent banners in Playwright
- How to handle popups and modals in Playwright
- How to take full-page screenshots with Playwright
- How to generate a PDF with Playwright and Firefox
- How to wait for content to load in Playwright
- How to retry failed requests when scraping Playwright
- How to scrape pages in parallel with Playwright
- How to rate limit your own Playwright scraper
- How to scrape HTML tables with Playwright
- How to scrape iframe content with Playwright
- How to scrape shadow DOM content with Playwright
- How to capture XHR and API responses in Playwright
- How to scrape geotargeted content with Playwright
- How to scrape real estate listings with Playwright
- How to scrape job postings with Playwright
- How to scrape e-commerce product pages with Playwright
- How to track product prices with Playwright
- How to scrape hotel room prices with Playwright
- How to scrape flight prices with Playwright
- How to scrape classifieds listings with Playwright
- How to scrape vacation rental listings with Playwright
- How to scrape car listings with Playwright
- How to scrape apartment rentals with Playwright
- How to track product stock and restocks with Playwright
- How to scrape location-based store prices with Playwright
- How to scrape flexible-date fare calendars with Playwright
- How to scrape product reviews with Playwright
- How to scrape reviews and ratings with Playwright
- How to scrape news article text with Playwright
- How to scrape business directory listings with Playwright
- How to scrape event and ticket listings with Playwright
- How to scrape restaurant menu data with Playwright
- How to scrape stock and financial data with Playwright
- How to scrape social media profiles with Playwright
- How to scrape forum and community threads with Playwright
- How to scrape image galleries with Playwright
- How to scrape video listings and metadata with Playwright
- How to scrape map-based local results with Playwright
- How to scrape sports scores and stats with Playwright
- How to scrape cryptocurrency prices with Playwright
- How to scrape deals and coupon codes with Playwright
- How to scrape to CSV with Playwright
- How to scrape to JSON Lines with Playwright
- How to scrape into a SQLite database with Playwright
- How to export scraped data to Excel with Playwright
- How to extract JSON-LD structured data with Playwright
- How to extract Open Graph and meta tags with Playwright
- How to extract links and build a crawl frontier in Playwright
- How to scrape RSS and Atom feeds with Playwright
- How to download images in bulk with Playwright
- How to extract clean article text with Playwright
- How to scrape a sitemap.xml with Playwright
- How to scrape into a pandas DataFrame with Playwright
- How to clean scraped prices and dates with Playwright
- Scrape search results by driving a form in Playwright
- Scrape a map-based search with Playwright
- Scrape autocomplete and typeahead inputs with Playwright
- Scrape date-picker calendars with Playwright
- Crawl list pages to detail pages with Playwright
- Scrape lazy-loaded images with Playwright
- Extract data from canvas charts with Playwright
- Scrape a multi-step wizard flow with Playwright
- How to resume an interrupted scrape with Playwright
- Incremental scraping: only new items since last run
- Handle 403 and 429 backoff mid-scrape in Playwright
- Scrape load-more button pages with Playwright
- Scrape nested pagination with Playwright
- Scrape an SPA that changes URL via history API
- Use BeautifulSoup with invisible_playwright
- Run stealth Playwright tests with pytest fixtures
- Run invisible_playwright concurrently with asyncio
- Run invisible_playwright in GitHub Actions CI
- Can you run invisible_playwright serverless?
- Run invisible_playwright in Celery task workers
- Schedule invisible_playwright scrapes with cron
- Run invisible_playwright headful on a server with Xvfb
- Use invisible_playwright in an Airflow DAG
- Combine invisible_playwright with httpx for speed
- Wrap invisible_playwright in a FastAPI service
- Run invisible_playwright in a Jupyter notebook
- Block images to speed up scraping (and when not to)
- Wait for a specific API response in Playwright
Comparisons
- Playwright stealth in Python: three levels that work
- Firefox or Chromium for anti-detect automation
- Chromium is not Chrome, and detectors know the difference
- Playwright stealth vs Camoufox: two patched Firefoxes
- Playwright stealth vs Patchright: driver vs engine
- Playwright stealth vs undetected-chromedriver and nodriver
- playwright-stealth vs a patched engine: page vs browser
- puppeteer-extra-plugin-stealth: unmaintained since 2024
- selenium-stealth hasn't been updated since December 2021
- pyppeteer's own maintainer says to switch to Playwright
- invisible_playwright vs rebrowser-patches: the same CDP fix
- invisible_playwright vs fingerprint-suite: injection vs engine
- invisible_playwright vs playwright-with-fingerprints
- invisible_playwright vs Scrapling
- invisible_playwright vs Ulixee Hero
- invisible_playwright vs SeleniumBase UC Mode
- Splash is unmaintained, and it was never a real browser
- invisible_playwright vs DrissionPage
- WebDriver BiDi vs CDP: does the new protocol hide you
- invisible_playwright vs hrequests
- zendriver vs invisible_playwright: Chrome CDP vs Firefox
- botasaurus vs invisible_playwright: framework vs library
- curl_cffi vs invisible_playwright: TLS client vs browser
- pydoll vs invisible_playwright: CDP without a driver
- selenium-driverless vs invisible_playwright stealth
- puppeteer-real-browser vs invisible_playwright
- Migrating from Selenium to Playwright for stealth
- Migrating from Puppeteer to Playwright for stealth
- undetected-chromedriver vs a patched Firefox browser
- scrapy-playwright vs a patched Firefox for stealth
- playwright-extra stealth plugins vs a patched browser
- tls-client vs a real browser: when TLS is enough
- Anti-detect browser or Playwright stealth: which you need
- undetected-playwright vs a patched Firefox binary
Integrations
- Using invisible_playwright with CodeceptJS
- Using invisible_playwright with Crawlee for Python
- Using invisible_playwright with Crawlee for JavaScript
- Using invisible_playwright with scrapy-playwright
- Using invisible_playwright with Robot Framework Browser
- Cypress, WebdriverIO, TestCafe and Nightwatch integration
- Using invisible_playwright with Microsoft's Playwright MCP
- Using the engine from Go, Java, C#, Ruby and Rust
docs/ source folder