botstop.io answers one question about whoever opens it: does this browser look like a bot? There is no CAPTCHA and nothing to click. The page takes its measurements, gives a verdict within about a second (human, suspicious or automated) and then shows every signal behind it: what it looked at, the value it found in your browser, and what that value means.
This repository holds no code. It explains what the site measures and how automated browsers still give themselves away, and it keeps the score of the scraping and AI-agent browser stacks that were sent there.
Try it on botstop.io · Scoreboard · Submit a screenshot · What it does · How bots still get caught · Further reading
29 scraping and screenshot APIs, cloud browsers for AI agents, and headless and anti-detect browsers, each sent to botstop.io like any other visitor. The number is the suspicion score the page gave it, from 0 to 100: the higher, the more plainly a bot. Last updated November 2026.
| # | Stack | Score | Verdict | Tested | Screenshot |
|---|---|---|---|---|---|
| 1 | Urlbox Website screenshot and rendering API |
100 | 2026-11-13 free plan real browser max stealth |
view | |
| 2 | Microlink API that turns a URL into data, a screenshot or a PDF, with real browsers |
100 | 2026-11-13 free plan real browser max stealth |
view | |
| 3 | Diffbot Extraction and crawling APIs, and a knowledge graph of the public web |
100 | 2026-11-13 free plan real browser max stealth |
view | |
| 4 | ScrapeNinja Web scraping API with Chrome fingerprint emulation and a real-browser engine |
100 | 2026-11-13 free plan real browser max stealth |
view | |
| 5 | Zenrows browserCloud-hosted browser for Puppeteer, Playwright or any other CDP client |
100 | 2026-11-13 free plan real browser max stealth |
view | |
| 6 | ScrapingAnt Web scraping API with headless Chrome and rotating proxies |
100 | 2026-11-13 free plan proxy headless browser |
view | |
| 7 | Decodo Proxy networks and a web data API; formerly Smartproxy |
100 | 2026-11-13 free plan real browser |
view Decodo's Markdown of the page, not a screenshot |
|
| 8 | ScrapingBee Web scraping API with headless Chrome and proxy rotation |
100 | 2026-11-13 free plan real browser |
view | |
| 9 | Hyperbrowser Cloud-hosted Chrome sessions and web data APIs for AI agents |
98 | 2026-11-13 free plan real browser max stealth |
view | |
| 10 | Firecrawl agentAgent that searches and navigates the web for the data you describe |
90 | 2026-11-13 free plan real browser max stealth |
view | |
| 11 | Browserless Managed headless browsers for Puppeteer, Playwright and Selenium |
85 | 2026-11-13 free plan default settings |
view | |
| 12 | Zyte API Web scraping API with automatic ban handling and browser rendering |
85 | 2026-11-13 free plan real browser |
view | |
| 13 | Anchor Browser Cloud-hosted Chromium browsers for AI agents |
82 | before 2026-11-13 free plan real browser max stealth |
pending | |
| 14 | ScrapeOps One endpoint in front of many scraping APIs and proxy providers |
70 | before 2026-11-13 free plan real browser max stealth |
pending | |
| 15 | Browser Use The open-source browser agent and its cloud of stealth browsers |
55 | 2026-11-13 free plan default settings |
view | |
| 16 | Simplescraper Point-and-click web scraper with cloud runs and an extraction API |
45 | 2026-11-13 free plan default settings |
view | |
| 17 | Firecrawl apiScraping API: a URL in, Markdown, HTML, a screenshot or JSON out |
40 | before 2026-11-13 free plan real browser max stealth |
pending | |
| 18 | Context.dev Web scraping API for Markdown, HTML, screenshots or JSON; formerly Brand.dev |
33 | 2026-11-13 free plan default settings |
view | |
| 19 | Scrape.do Web scraping API with rotating proxies and JavaScript rendering |
33 | 2026-11-13 free plan real browser |
view | |
| 20 | Lightpanda Headless browser written from scratch in Zig, without graphical rendering |
27 | 2026-11-13 free plan default settings |
view Lightpanda's own render: it draws no pixels |
|
| 21 | Kernel Isolated Chromium sessions in the cloud for AI agents |
25 | 2026-11-13 free plan headful, stealth no GPU acceleration |
view | |
| 22 | Scrapfly cloud browserCloud-hosted browsers for CDP clients, with an anti-bot Unblock Mode |
25 | 2026-11-13 free plan Unblock Mode on |
view | |
| 23 | Scrapfly screenshot apiScreenshot API: a URL in, an image of the rendered page out |
25 | 2026-11-13 free plan real browser |
view | |
| 24 | Scrapfly api, chromeWeb scraping API with browser rendering, here presenting itself as Chrome |
25 | before 2026-11-13 free plan real browser max stealth |
pending | |
| 25 | Scrapfly api, operaThe same API with the browser presenting itself as Opera |
25 | before 2026-11-13 free plan real browser max stealth |
pending | |
| 26 | Zenrows fetchScraping API with JavaScript rendering, proxy rotation and anti-bot handling |
25 | before 2026-11-13 free plan real browser max stealth |
pending | |
| 27 | Crawlbase Crawling API with proxies, browser rendering and anti-bot handling built in |
15 | before 2026-11-13 free plan real browser max stealth |
pending | |
| 28 | Camoufox headlessOpen-source anti-detect Firefox that changes its fingerprint in C++ |
15 | before 2026-11-13 free plan real browser max stealth |
pending | |
| 29 | Camoufox headful, xvfbThe same browser with a window, on a virtual display (Xvfb) |
0 | before 2026-11-13 free plan real browser max stealth |
pending |
How to read it:
- Sixteen of the 29 are called
automated, tensuspicious, and three pass ashuman, one of them without a single point. The notches on each bar are the two thresholds, 18 and 45. - A score belongs to one product in one configuration, on the day it was tested. The same vendor lands in different places with different products: Firecrawl's agent scores 90 and its API 40, Zenrows' Browser 100 and its Fetch API 25. Running Camoufox headless costs it 15 points; with a window on a virtual display it scores 0.
- The detector changes too. Every check carries the detector version it was made with, shown in the header of the page.
- The IP address plays no part. The score is made of what the browser and its machine give away. A datacenter address costs nothing here, and a residential proxy gains nothing.
- Much of what is counted is specific to Chrome and the DevTools protocol, as the tables below show. Camoufox is a Firefox build and is driven over a different protocol.
Tested a stack yourself? Open a scoreboard issue with its screenshot, the score and the setup it ran with. Stacks already on the board are welcome again, with other settings or on a newer detector version.
- Send the stack to https://botstop.io and let the check finish: the result card says "all tests done", and the page adds the element
#allTestsDoneat that moment, so a screenshot API can wait for it. - Take the screenshot, full page if the stack allows it.
- Note the plan and the setup: real browser or plain HTTP, headless or headful, stealth and proxy settings. The form asks for each.
The check runs by itself when the page opens, in two phases: an instant verdict from the signals that are there immediately, refined once by the probes that need time.
sequenceDiagram
autonumber
participant B as Browser
participant S as botstop.io
B->>S: GET /
S-->>B: the page, the collector script, a single-use nonce
Note over B: the collector runs its probes
B->>S: sealed snapshot, well under a second in
S-->>B: verdict, score and every signal
Note over B: the Worker cross-check and the slow probes settle
B->>S: fuller sealed snapshot, at about 2.6 s
S-->>B: refined verdict
The score. A classifier on the server adds up weights. Every signal that fires contributes its points to a suspicion score from 0 to 100, and two thresholds turn the score into a verdict: 18 points make a visitor suspicious, 45 points make it automated. A strong signal carries 45 points, so one of them is enough. Strong signals are the ones in which the browser gives itself away, such as navigator.webdriver, a headless User-Agent or brand, or input in which every single event was synthetic.
Weight 0. A signal can also be measured, shown and not counted. The page labels it NOT SCORED. This is how a new signal is tried on real visitors before it is allowed to change anybody's verdict, and several of the most interesting ones below are still in that state.
Passive. The page asks you to do nothing, so sitting still is not held against you. Only positive evidence counts, and not touching the mouse costs nothing.
Everything is shown. Below the verdict the page lists all of its signals, 38 in five groups as of v13.0, each with the value it observed, a paragraph on what that value means, and the points it added. A panel called "Hardware Behind the Browser" puts what the browser claims next to what its machine was measured to be. Every check is stored and gets a permalink.
Versions. The header carries the version of the detector: the schema of the collector (what is measured), then the revision of the classifier (how it is judged). Every check keeps the version it was made with, so a permalink to an older check says what its version scored and what the current one makes of the same data.
Proxy and VPN, on request. The box under the verdict can also run the proxydetect.live engine on your connection, for a proxy score and a VPN score. It does so only when asked, with its button or with ?proxy=1 in the address, and it is a separate test: the bot verdict does not depend on it.
botstop.io is a thin page around a detector that also guards the sign-up and login forms of ipapi.is. That is where its signals are measured on real people before they are allowed to count.
What signs up to a web service today is rarely a script that speaks HTTP. It is a real Chrome on a rented server, steered by a program over the DevTools protocol (CDP), usually with Puppeteer or Playwright or something built on them, and usually behind a proxy. It runs a page's JavaScript like any other browser, and its operator controls everything the browser says about itself: the User-Agent, navigator.webdriver, the WebGL renderer string, the number of cores.
What the operator does not control as easily is what the browser does while a program drives it, and the machine it runs on. botstop.io asks five questions. The points are those of v13.0; the page always shows the current ones.
A page cannot see the DevTools protocol. It can notice that the browser behaves differently while a client is attached, or because of how it was started. On the page: Automation & CDP Detection.
| Signal | What the page sees | Why a bot shows it | Points |
|---|---|---|---|
| WebDriver flag | navigator.webdriver === true |
Selenium, Puppeteer and Playwright set it unless it is stripped | 45, strong |
| Headless or crawler User-Agent | HeadlessChrome, PhantomJS, Electron, Puppeteer or Playwright in the string, or a self-declared bot or crawler |
The browser names itself | 45, strong |
| Headless brand | HeadlessChrome is still in navigator.userAgentData.brands and in the Sec-CH-UA request header behind a replaced User-Agent string |
--user-agent replaces the string and nothing else |
45, strong |
| Replaced User-Agent | No brands at all in userAgentData, or a request without Accept-Language |
Puppeteer's page.setUserAgent() replaces the User-Agent over CDP without its metadata; the headless shell sends no language unless one is set |
18 each |
| DevTools protocol client | An error thrown by code compiled from a string reaches window.onerror with an empty source URL, and console.debug(new Error()) runs the page's Error.prepareStackTrace |
Both are side effects of Runtime.enable, which almost every automation library sends. Checked again in a fresh Worker, where page-level patches do not reach |
30, counted once |
| Mouse press without pressure | A trusted pointerdown from a mouse with a pressure of 0 |
A press sent through CDP's Input.dispatchMouseEvent, as Puppeteer does it. A real mouse reports 0.5 |
30 |
| History state rate limit | The 201st history.replaceState() within 10 seconds is accepted |
Chrome drops it. Puppeteer and Playwright start Chrome with --disable-ipc-flooding-protection, which lifts the limit |
30 |
| Patched native functions | Function.prototype.toString, property descriptors on built-ins, Proxy frames in the stack of a TypeError |
A stealth patch is itself visible | 18 |
| Notification permission | Notification.permission says denied while permissions.query() says prompt |
Unpatched headless Chrome | 18 |
window.chrome |
Missing behind a Chrome User-Agent | Old headless builds and User-Agent spoofers | 18 |
An attached DevTools client is not proof of abuse. Undocked DevTools look the same, and so does an AI agent that drives a person's own browser. That is why the client alone makes a visitor suspicious and never automated.
A browser on a server is started without the things a desktop has, and with switches that take pieces out of Chrome. On the page: Window, Viewport & Rendering Integrity.
| Signal | What the page sees | Why a bot shows it | Points |
|---|---|---|---|
| Window without browser chrome | The viewport fills the window: outerHeight against innerHeight, with page zoom taken out first |
Headless Chrome, Puppeteer and Playwright have no tab strip and no address bar above the page | 25 |
| Zero-size window | outerWidth or outerHeight of 0 on a visible page |
A frame with no window of the operating system behind it | 15 |
| No pointing device | (any-pointer: none) on a desktop |
Headless Chrome on Linux has no input devices. Headed Chrome on the same server has a mouse, even on a virtual display | 15 |
| Default screen | A screen of exactly 800×600 | Headless Chrome's screen, unless the operator sets another | 10 |
| No plugins | navigator.plugins is empty and there are no touch points |
A stripped headless configuration | 12 |
| No speech voices | Zero voices behind a macOS or Windows User-Agent | A Linux container dressed up as a desktop | 15 |
| Rounded heap limit | performance.memory.jsHeapSizeLimit is rounded to three digits |
Chrome hides the exact value where site isolation is off: the headless shell, or --disable-site-isolation-trials |
18 |
| Chrome's own URL schemes | new URL('chrome-extension://x:0') keeps its port |
chrome-headless-shell does not register the scheme |
18 |
| Widevine | Google Chrome without the Widevine key system | Playwright's --disable-component-update |
18 |
| Work-area clamping | Firefox on Windows whose window is exactly the size of the screen's work area, borders included | Anti-detect browsers that clamp the window to the work area, as Camoufox does | 45, strong |
Recorded and shown, but not scored: the page's icon is never fetched (the headless shell has no tab to show it in), Chrome's built-in language detector is unavailable, none of the five permissions that Chrome grants by default is granted, Google Chrome has no "Google …" speech voices, the keyboard has an IntlYen key on a non-Japanese Linux, a frame's message arrives before its load event, and DOM text is laid out at another width than canvas measures. Each of them gives away some automated stack in the lab. Each also fires for some real browsers, adds nothing to a signal that already counts, or has not been measured on enough of them.
Stealth patches are applied to the page. A dedicated Worker is a second JavaScript realm with a navigator of its own, and the patches usually do not reach it. The page starts one and compares. On the page: Worker Realm Cross-Check.
| Compared | Between the page and the Worker | Points |
|---|---|---|
| Identity | navigator.userAgent, navigator.platform, and the platform, architecture and bitness from userAgentData |
18 |
| Environment | hardwareConcurrency, deviceMemory, the language, the number of languages and the time zone |
30 |
| GPU | The WebGL renderer string, read in the Worker through an OffscreenCanvas |
30 |
Disagreements with one root cause count once, 30 points at most. A user-agent switcher extension and the device emulation of DevTools split the two realms in the same way, so a disagreement alone is suspicious.
A browser in a datacenter usually has no GPU. Chrome then renders with SwiftShader, a software rasterizer, and says so in the WebGL renderer string. So the string gets overwritten, and the server becomes an "NVIDIA GeForce RTX 3060" or an "Apple M4 Max". These probes do not ask the browser what it runs on. They measure it. On the page: Computing Environment & Side-Channels.
| Probe | What it measures | What it tells apart | Points |
|---|---|---|---|
| Renderer string | The WebGL vendor and renderer | SwiftShader, llvmpipe, lavapipe and the adapters of VirtualBox and VMware, or no WebGL at all | 15 |
| Software WebGL despite the flag | A context requested with failIfMajorPerformanceCaveat: true |
Headed Chrome refuses software WebGL. Headless Chrome hands out SwiftShader anyway | 20 |
| WebGPU against WebGL | The adapter's vendor, architecture, backend and isFallbackAdapter |
A "hardware" GPU on a fallback adapter; Vulkan on a claimed Mac | 30 |
| Shader time | Six blended passes of a fragment shader in which every pixel makes 96 texture lookups, each depending on the one before | On real sign-ins an Apple Silicon Mac needed 1 to 11 ms and every other desktop GPU at most 71 ms; software rasterizers needed 36 ms to over a second. Counted above 100 ms behind a hardware GPU string, above 30 ms behind an Apple Silicon one | 25 |
| Rasterizer precision | How finely the rasterizer places the edge of a triangle | A GPU snaps vertices to 1/256 px, SwiftShader to 1/16 px. It is the shader finding without a clock, and the two count once | 25 |
| CPU architecture and L1 cache | The sign of the NaN that Infinity - Infinity leaves in a Float32Array, and a pointer chase through 96 KB against one through 16 KB |
The sign bit is set on x86 and clear on ARM. The two chases take equally long where 96 KB fit the L1 data cache, as on Apple's performance cores, and about twice as long on x86. Counted for an "Apple Silicon Mac" that has both wrong | 25 |
| CPU instruction set | Eight WebAssembly relaxed-SIMD operations whose result the specification leaves to the CPU | x86 or ARM, with or without FMA, and no clock involved. Counted for a phone on x86, for client hints that name x86 on an ARM CPU, and for an Apple Silicon Mac on x86 with FMA, which Rosetta does not have | 30 |
Recorded and shown, but not scored: what arithmetic on subnormal numbers costs (Intel CPUs hand it to microcode and take 17 to 65 times as long, AMD and Apple do not), and how many page faults 16 MB of new memory take (Apple Silicon has 16 KB pages, Linux and Windows fault 4 KB at a time).
What that buys: one fleet that visits botstop.io presents itself as a MacBook Pro with an M4 Max, down to a matching WebGPU adapter and client hints. Its shader took between 36 and 208 ms in every one of 121 visits. Among 271 sign-ins from real Apple Silicon Macs, none needed more than 11.1 ms.
None of this is unbeatable. Whatever a probe reads can be replaced, by a script in the page or in the C++ of an anti-detect browser. The probes raise the price: the cheapest way to pass them is to own the hardware one claims to have.
On the page: Behavioral & Anti-Replay Telemetry.
- Synthetic input. The collector watches 22 kinds of DOM events and reads
event.isTrusted. Events that a script dispatched are untrusted. A visitor whose every event was synthetic, pointer movement included, gets a strong signal (45 points); a large share of synthetic events with no pointer activity counts 20. - Honeypot. An input placed off screen, where no person sees it. Form fillers that populate every
<input>fill it too (30 points). - Anti-replay challenge. Every page load gets a single-use nonce. The collector signs it, together with the geometry of the environment and a digest of the probe results, inside a small bytecode VM that is randomized with every build, and seals the snapshot with X25519 and XSalsa20-Poly1305. A snapshot can therefore not be replayed, or edited on the way, over plain HTTP.
A detector is only as good as its false positives are rare. Most of the signals above were counted on real traffic before they got a weight, and a good part of what was tried did not make it:
- No interaction. Not moving the mouse says nothing on a page that asks for nothing.
- Timing a CPU loop, Worker round trips or Worker scaling. Meant to expose hypervisors and overclaimed core counts. On the sign-in pages of ipapi.is the round trips fired for 25% of paying customers: they measured the page's own start-up work.
- A frame's message that overtakes its
loadevent. Stable in every lab run with Puppeteer and Playwright. On real traffic it fires for half of the paying customers. - A pointer press without pressure, from any pointer. It is what CDP produces, and what every finger on an iPhone produces. It now counts for a mouse only.
Open https://botstop.io in the browser you want judged, or send your automation there and read the result card, or take a screenshot as above. Give it three seconds: the first verdict appears at once and is refined when the slower probes have settled, at which point the card says "full scan complete · all tests done" and the page adds the element #allTestsDone, the selector to wait for before a screenshot. A client that does not run JavaScript gets no verdict, because the check is the script.
botstop.io/?proxy=1 runs the proxy and VPN tests after the bot check.
If you get a stack past it, or it calls your real browser a bot, open an issue here and include the permalink of the check. To put a stack on the scoreboard, submit a screenshot.
- The Machine Behind the Browser: New Chrome Automation Signals and Timing Side Channels on botstop.io (November 2026). Where the signals come from, with code for the main probes, and what eleven days of real sign-ins did to them.
- botstop.io: Does your Browser look like a Bot? (August 2026). The launch post.
- Browser Red Pills: Why are you browsing my website from AWS Lambda? (2021). The unfinished post that the hardware probes grew out of.
- On the Architecture of Bot Detection Services and Behavioral Analysis for Bot Detection (2021). The theory behind it.
Built by Nikolai Tschacher. Detection by the ipapi.is behavior engine, proxy and VPN tests by proxydetect.live.