Anti-Detect Browser Tools — Technical Comparison

Source-verified comparison of ten open-source browser-automation and anti-detection tools. Each claim carries an evidence tier stating how it was established.

• Interactive view: docs/index.html — filterable matrix, findings, metrics, lineage
• Verification date: 2026-10-13
• Version and project-health data: STATUS.md, generated by python scripts/verify.py --write
• Code quality, security patterns, maintainability: CODE-REVIEW.md
• Method, scope, and limits: METHODOLOGY.md
• Revision history and corrections: CHANGELOG.md

Disclosure

[!IMPORTANT]

This repository is sponsored by Scrappey, a commercial

web-data API — a paid substitute for the tools compared here. The sponsor has a

commercial interest in how these tools are described.

Constraints applied as a result:

• Scrappey is not rated, scored, or listed in any comparison table in this repository.
• Claims about Scrappey are subject to the same evidence tiers as any other vendor. No

• Disagreements about a specific claim are handled as defects: open an issue

Full policy: METHODOLOGY.md § Conflict of interest.

Evidence tiers

| Tier | Definition | Established by | |:----:|-----------|----------------| | A | Verified in source at a stated commit | Reading the code in an isolated container; file paths cited | | B | Vendor-reported, not reproduced here | The project's README, release notes, or docs | | C | Community-reported | Issue trackers, forums, third-party posts | | D | Not established | No evidence located in either direction |

Two constraints follow from the method (see METHODOLOGY.md):

1. No claim about anti-bot service outcomes in this repository exceeds Tier B. No

1. Vendor-published benchmarks remain Tier B regardless of detail. They record what

[!NOTE]

Tier D means nobody published a test — not that the tool failed one. Collapsing

"not established" into "does not implement" is the most common way a comparison

misleads, so the two are never merged here.

Tools covered

| Tool | Implementation | Language | License | Page | |------|---------------|----------|---------|------| | Camoufox | Firefox fork; fingerprint controls compiled into C++ | Python; TypeScript launcher in the repository | MPL-2.0 (browser), MIT (wrapper) | camoufox.md | | Patchright | Playwright driver source patched at build time via AST rewrite | Python, Node, .NET | Apache-2.0 | patchright.md | | SeleniumBase | Patched ChromeDriver (UC Mode) + CDP-native driver (CDP Mode) | Python | MIT | seleniumbase.md | | Botasaurus | Raw-CDP driver + scraping framework | Python | MIT | botasaurus.md | | XDriver | File-replacement patcher shipping a prebuilt playwright-core fork | Python | Apache-2.0 | xdriver.md | | CloakBrowser | Chromium fork with C++ patches; open wrapper, closed binary | Python, Node, .NET | MIT (wrapper), proprietary (binary) | cloakbrowser.md | | Scrapling | Scraping framework: HTTP and browser tiers, parser, spider | Python | BSD-3-Clause | scrapling.md | | Obscura | Browser engine implemented from scratch in Rust | Rust; CDP clients | Apache-2.0 | obscura.md | | Clearcote | ungoogled-chromium fork; published patches (open build), fingerprint import | Python, Node, .NET | BSD-3-Clause | clearcote.md | | invisible_playwright | Firefox fork patched in C++, driven through a vendored, modified copy of Playwright's Python client and an in-process Python Juggler server | Python | MIT AND Apache-2.0 (wrapper, as declared in pyproject.toml; the vendored client is Apache-2.0), MPL-2.0 (engine fork) | invisible-playwright.md |

Versions, release dates, star counts, contributor counts, and last-push dates: STATUS.md.

Changes since the 2026-09-30 revision

Seven of ten tools published releases; the full list, with corrections, is in CHANGELOG.md. Material changes to previously documented behaviour:

• Camoufox (v156.0.1-beta.38, PyPI 0.5.8, still Firefox 156.0.1): the repository

• invisible_playwright (0.25.7 to 0.35.3; core 34.31.0 to 41.39.0; engine

• CloakBrowser 0.5.11 to 0.6.2: the Pro binary is Chromium 154 on every platform,

• Clearcote: no engine change (v0.1.0-pre.23, Chromium 150, 37 patches). SDKs 0.33.0

• Obscura v0.2.3 to v0.2.4: intercepted XHR is reported as Xhr rather than

• SeleniumBase 4.54.13 to 4.56.1: an opt-in Fast Mode that skips built-in waits

• Patchright (1.63.0), Scrapling (0.4.15), Botasaurus and XDriver: no new

• Code metrics were regenerated. scripts/codemetrics.py now also excludes the

Corrections to claims that were already wrong at the 2026-08-14 commits, and to claims that were unsupported rather than merely stale, are itemised in CHANGELOG.md.

Architecture

Tier A unless a cell states otherwise. Read from source at the commits listed in STATUS.md. Cells state what the code does; they are not ratings.

Stealth implementation

| | Camoufox | Patchright | SeleniumBase | Botasaurus | XDriver | CloakBrowser | Scrapling | Obscura | Clearcote | invisible_playwright | |---|---|---|---|---|---|---|---|---|---|---| | Browser engine | Firefox 156.0.1 fork | stock Chromium | stock Chromium | stock Chromium | stock Chromium | Chromium 152 fork (Pro; 151 on macOS) / 146 (keyless free) — Tier B | delegates | own Rust engine | Chromium 150 fork (ungoogled; open build) | Firefox 151 fork | | Modification method | C++ source patches: 42 in browser/patches/ (31 diff C/C++ sources), 2 in browser/patches/playwright/ | AST rewrite of the Playwright driver (patchright_driver_patch.ts, 275 lines, ts-morph) | ChromeDriver binary rewrite (undetected/patcher.py) + CDP driver (undetected/cdp_driver/) | raw CDP over WebSocket; 58 generated binding files in botasaurus_driver/cdp/ | file replacement with a prebuilt bundle (turnstilebrowser-playwright-core 1.49.0) | C++ source patches: 87 (Pro), 58 (keyless free), 26 (macOS keyless free) — Tier B, counts from vendor README | dependency on Patchright + curl_cffi | JS shims + native ops in its own runtime | C++ source patches: 37, listed in patches/series | C++ source patches: 104 files on a public Firefox fork; no patch series | | navigator.webdriver | C++ patch | driver patch | driver patch | CDP | prebuilt bundle | C++ patch (Tier B) | inherited from Patchright | runtime shim | C++ patch 010-user-agent-and-webdriver | C++ patch | | Runtime.enable tell | not applicable — automation runs over Juggler, not CDP; the only reference is a comment in a build-checker (browser/tests/build-tester/src/lib/checks/core.ts:56) | Runtime.enable sends are removed in the driver patch (crPagePatch.ts, crDevToolsPatch.ts, crServiceWorkerPatch.ts); isolated contexts and a disabled Console API are described in the project README (Tier B) | Tier D — no handling located in source; CDP Mode attaches no WebDriver | none — the sole reference is a commented-out line, connection.py:219 | rebrowser-lineage patch in the bundle, gated behind TURNSTILEBROWSER_PATCHES_* env vars | Tier D — not addressed in the wrapper README's patch list; binary is closed and cannot be inspected | inherited from Patchright | own CDP server implementation | C++ patch 110-runtime-enable | not applicable — Juggler, not CDP | | TLS/JA3 impersonation | none — real Firefox TLS | none — real Chrome TLS | none | none | none | none — real Chrome TLS | yes, HTTP tier only: curl_cffi>=0.16.1 | optional, build-gated: wreq behind --features stealth (only the -stealth release archives) | stock Chrome TLS by default; patch 210-tls-network-persona swaps the post-quantum key-share group and the ALPS codepoint only when the claimed Chrome major differs from the build's | none — real Firefox TLS | | Input simulation | humanize, humanize:minTime, humanize:maxTime config properties (browser/settings/properties.json:41-43); the path replays recorded human movements (Cursory, 2,356 recordings, vendored cursory-js, LGPLv3-or-later) generated in Juggler JS (browser/additions/juggler/input/CursorTrajectory.js), not in C++ | none; the project pairs with the separate CDP-Patches library, whose own README (v1.1, 2025-09-28) says the Chromium input leak is fixed upstream and the package is still needed only for Select elements (Tier B) | UC Mode: PyAutoGUI moveTo(..., 0.25, easeOutQuad) plus click (browser_launcher.py:1257-1284); CDP Mode click(): Input.dispatchMouseEvent move, then press and release at the element centre ±0.875 px with randomised pauses (cdp_driver/element.py:498-583); no motion model | Bézier trajectories with Gaussian distortion and easing (human_curve_generator.py) | none — the only matches are incidental strings in bundled Playwright assets | humanize option: Bézier mouse paths, typing with typos, scroll physics, DOM reads in an isolated world (wrapper source, Tier A); the binary-side CDP input mimicking is Tier B (closed binary) | none first-party; passes options through to Patchright (only the Cloudflare solver randomises its click offset and press delay, _stealth.py:170-173) | none — bezier matches are CSS cubic-bezier easing and canvas bezierCurveTo | C++ patches 130-humanized-input, 200-agent-and-humanized-input; the SDK also builds eased Bézier paths and sends them through Playwright's native mouse; recorded human trajectories are licensed-build only (project README, Tier B) | client-side: a seeded path planned in Python (invisible_core.juggler._motion, wrapper _cursor.py), on by default (humanize=True, launcher.py:104) (Tier A) | | Engine source published | yes — patches + build script; binary is rebuildable | n/a — patches stock Chromium | n/a — patches stock ChromeDriver | n/a — drives stock Chromium | no — prebuilt bundle; upstream rebrowser-patches is public, this bundle is not built from published sources | no — no patches directory or build script in the repository; binary is proprietary | n/a — delegates | yes — the repository is the engine | yes (open build) — patches + build docs; GPG-signed, SHA-256-checksummed releases. Licensed builds carry private patches and are not reproducible from public source (project README, Tier B) | yes — full public Firefox fork, but no patch series |

Non-stealth capabilities

| | Camoufox | Patchright | SeleniumBase | Botasaurus | XDriver | CloakBrowser | Scrapling | Obscura | Clearcote | invisible_playwright | |---|---|---|---|---|---|---|---|---|---|---| | Layout and rendering | Gecko | Blink | Blink | Blink | Blink | Blink | Blink (browser tier) | own engine, 72,401 lines — render-enabled builds only | Blink | Gecko | | HTML parser included | no | no | no | no | no | no | yes | yes (html5ever) | no | no | | Crawler framework | no | no | no | yes | no | no | yes | no | no | no | | CAPTCHA handling | no | no | click-based solving for several vendors | Cloudflare challenge automation | no | no | solve_cloudflare for Turnstile/Interstitial | no | no | no | | Test framework integration | no | no | pytest, unittest | no | no | no | no | no | no | no |

Notes on reading these tables

• Patch counts measure surface area touched, not quality. Projects split changes at

• "Delegates" describes a dependency relationship, not a deficiency. Scrapling's

• Absence of TLS impersonation is expected for real-browser tools. A real browser

• Two cells changed direction during verification on 2026-08-14: SeleniumBase and

• **Several cells were corrected on 2026-09-30 for claims that were already inaccurate at

Anti-bot service claims

[!WARNING]

No benchmark was run for this table. It records who claims what, not what works.

Vendor checkmarks are not comparable across projects — each used undisclosed targets,

dates, IP types, and pass criteria.

What follows is the provenance of each project's own claims.

| Tool | Services the project claims | Tier | Basis stated by the project | |------|---------------------------|:----:|------------------------------| | Patchright | Cloudflare, Kasada, Akamai, DataDome, Shape/F5, Bet365, Fingerprint.com, CreepJS, and others | B | README list; no methodology, targets, dates, or trial counts published | | XDriver | Cloudflare WAF, Turnstile, DataDome, Kasada, PerimeterX, Imperva, Fingerprint.com | B | README table dated September 2025. The bundle is pinned to playwright-core 1.49.0 and the repository has no commits since 2025-09-10 | | Botasaurus | Cloudflare WAF, Cloudflare Turnstile, DataDome | B | Test functions exist in bot_detection_tests.py (Tier A) targeting vendor demo endpoints; pass results are the vendor's claim | | CloakBrowser | Turnstile, FingerprintJS, reCAPTCHA v3 score 0.9, "30+ detection sites" | B | README results table; the strongest figures are gated to the paid Pro binary; the keyless free binary is not claimed to reach them. A free GitHub-sign-in key tier (one concurrent session) receives the latest build | | SeleniumBase | Cloudflare, Imperva/Incapsula, DataDome, Kasada, PerimeterX, reCAPTCHA | B | Runnable example scripts committed to examples/cdp_mode/ (their existence is Tier A); outcomes are not published as measurements | | Scrapling | Cloudflare Turnstile | B | solve_cloudflare exists in source (Tier A). The README's sponsor table lists a third-party paid token API for Akamai, DataDome, Kasada, and Incapsula rather than claiming them | | Camoufox | none | D | Makes no service claims. Its README documents a limitation: some WAFs probe SpiderMonkey engine behaviour, which a Firefox fork does not conceal (Tier B) | | Obscura | none | D | No service claims in the README; its docs list Cloudflare interactive challenges, DataDome and Akamai active challenges, and CAPTCHAs as not handled (Tier B) | | Clearcote | none | D | Publishes open-auditor results (CreepJS) and a self-referential coherence gate; no WAF pass rates | | invisible_playwright | none | D | Makes no WAF claims. Publishes detector-suite results (CreepJS, BotD, FingerprintJS, fpscanner, Sannysoft, BrowserLeaks, reCAPTCHA) with a public harness — Tier B for those results |

Constraints on interpreting this table

• It records claims, not capability. A claim's presence is evidence about the project's

• Vendor checkmarks are not comparable across projects. Each used undisclosed targets,

• The previous revision of this document contained a 7-service × 9-tool grid of

• Claim breadth in this table does not correlate with maintenance status. XDriver states

Lineage — ten tools, fewer independent implementations

Several of these depend on each other. Selecting two for redundancy can mean selecting the same implementation twice: if that lineage is fingerprinted, both fail together. Every edge below was verified in source (Tier A).

graph LR
  subgraph FF["Firefox lineage"]
    direction LR
    Firefox["Firefox 156.0.1"] --> Camoufox
    FPGen["fpgen<br/><i>fingerprint generator, repo HEAD</i>"] --> Camoufox
    Firefox151["Firefox 151"] --> FAP["firefox_antidetect_patch<br/><i>public fork, 104 files</i>"]
    FAP --> InvisiblePW["invisible_playwright"]
    PWClient["Playwright Python client<br/><i>vendored, modified</i>"] --> InvisiblePW
  end
subgraph CR["Chromium fork lineage"]
    direction LR
    Chromium --> CloakBrowser["CloakBrowser<br/><i>closed binary</i>"]
    Chromium --> UG["ungoogled-chromium"] --> Clearcote
  end
subgraph PW["Playwright-patch lineage"]
    direction LR
    Playwright --> Patchright --> Scrapling
    CDPPatches["CDP-Patches<br/><i>input</i>"] -.pairs with.-> Patchright
    Playwright --> Rebrowser["rebrowser-patches"]
    Rebrowser --> TB["turnstilebrowser-<br/>playwright-core 1.49"] --> XDriver
    Rebrowser --> RPC["rebrowser-playwright-core"] --> BotaJS["Botasaurus JS driver"]
  end
subgraph CDP["Raw-CDP lineage"]
    direction LR
    NoDriver -.derived.-> SeleniumBase["SeleniumBase<br/><i>CDP Mode</i>"]
    Botasaurus["Botasaurus<br/><i>own CDP driver</i>"]
  end
subgraph OWN["Independent engine"]
    direction LR
    V8["V8 / deno_core"] --> Obscura
    H5["html5ever"] --> Obscura
  end

What this shows:

• Scrapling's browser tier is Patchright (patchright>=1.62.1; the latest Patchright is

• SeleniumBase's CDP Mode is NoDriver-derived — stated in the file headers of

• XDriver and Botasaurus's JS driver both descend from rebrowser-patches

• Clearcote's CREDITS.md names two upstream projects — fingerprint-chromium as the

• Camoufox, Obscura, Clearcote and CloakBrowser are genuinely independent of the

• invisible_playwright is a second, independent Firefox lineage — its own C++ fork,

Code quality

Detection capability and code quality are independent properties. Full analysis — size, typing, tests, error handling, dependency hygiene, security patterns, with per-file citations — is in CODE-REVIEW.md. Metrics are reproducible via scripts/codemetrics.py.

Findings that bear on a build decision:

| Finding | Detail | |---------|--------| | CloakBrowser disables the Chromium sandbox by default, and Playwright's chromium_sandbox=True cannot restore it | --no-sandbox is part of the wrapper's default arguments (config.py:155-177, literal at :165; the Node and .NET wrappers and cloakserve do the same). Playwright itself adds the flag unless chromiumSandbox is true, but the wrapper adds it as well, so that switch has no effect under the default arguments. The argument merge can add and override flags but cannot remove them; the only opt-out is stealth_args=False, which also drops the wrapper's fingerprint seed and platform flags (the README shows re-adding --fingerprint by hand, README.md:822-825). Combined with a closed binary, this means unauditable native code processing hostile input with no OS sandbox. The README does not mention the flag; the CHANGELOG does. | | XDriver ships a frozen bundle with no drift detection | A prebuilt playwright-core 1.49.0 fork, 295 lines of first-party Python, zero tests. Nothing detects that upstream has moved. | | Patchright handles a fragile approach well | AST patching that fails loudly (225 …OrThrow( calls across driver_patches/) plus check_patch_impact.yml, which diffs Playwright versions and writes an impact report; in the release workflow a failed test run opens an issue and a draft PR. | | SeleniumBase is large and largely untyped | 102,429 production lines; base_case.py is 17,703 lines and 579 methods in one class; 3.4% of functions carry any annotation; 514 broad except handlers; 551 sleep calls. Forking or debugging it is correspondingly expensive. Its 42 shell=True calls include 12 in core/ launch helpers. | | Obscura's panic containment is partial | 95 matching lines of non-test .unwrap() and 1 panic! across 179,727 production lines (inline #[cfg(test)] modules stripped; the earlier 2,326 figure included test code). Release builds use panic = "unwind", each CDP connection runs on its own OS thread, and catch_unwind guards the JS-op, URL-op, module-evaluation and resource-load boundaries. Stack overflows, allocation failures and V8 fatal errors still abort the process. Relevant if you run obscura serve as a long-lived service. | | Botasaurus swallows errors | 38 bare except: clauses; botasaurus-driver has zero test files across 50,169 lines; PyPI (4.0.101) is ahead of the published source (4.0.92). | | Scrapling pairs high typing coverage with a low smell count | 85.5% annotation coverage, 60.4% docstrings, 0 bare excepts, 7 broad excepts across 16,219 lines, 63 test files, the only pre-commit config of the ten. |

Project health

Counts are not duplicated here. Stars, forks, contributors, open issues, last-push dates, and published versions live in STATUS.md, which is generated by scripts/verify.py. Open-issue counts in particular change within hours, so a hand-copied figure is wrong shortly after it is written; STATUS.md carries its own generation date.

Stable, decision-relevant properties as of 2026-10-13:

| Property | Finding | |----------|---------| | Proprietary components | CloakBrowser: its MIT license covers the Python/Node/.NET wrapper; the Chromium binary is proprietary, and the keyless free and Pro builds differ in Chromium version and patch count. Clearcote: the open build is BSD-3-Clause end to end; its licensed builds (free-with-GitHub key, Pro) are key-gated and carry private patches that are not in the public tree (project README, Tier B; their licence terms were not examined). Every other tool is permissively licensed end to end. | | Split licensing | Camoufox: MPL-2.0 for the Firefox fork, MIT for the Python wrapper; the vendored Cursory trajectories (browser/additions/juggler/input/cursory) are LGPLv3-or-later. invisible_playwright: MPL-2.0 engine fork; the wrapper and, since 38.34.0, invisible_core declare MIT AND Apache-2.0 (the vendored Playwright and Juggler client code is Apache-2.0). Relevant if you redistribute the browser. | | Maintenance concentration | Clearcote has 1 contributor, XDriver 2 and invisible_playwright 8. The other seven have 9 to 81. | | Inactive project | XDriver, last push 2025-09-10 — approximately 13 months before the verification date. Botasaurus's monorepo was last pushed on 2026-07-26, about two and a half months before it. The other eight were pushed within the 2 days preceding it. | | Registry ahead of source | botasaurus-driver publishes 4.0.101 to PyPI while the repository's setup.py reads 4.0.92. The installed artifact does not correspond to a tagged commit. | | Startup network traffic | Browser-originated launch request: invisible_playwright only. Its browser GETs a GitHub asset on every launch, on by default (invisible_firefox.usage_ping.enabled, firefox.js:3683; the core sets the URL pref but never enabled). It is disclosed in the engine README and in the wrapper's CHANGELOG.md:1346-1361, and the wrapper README shows a launches badge (README.md:309), but the wrapper README text and docs do not describe it. Whether the request leaves through the session proxy was not run; source order suggests the proxy command arrives later in startup (Tier D). Disable with InvisiblePlaywright(extra_prefs={"invisible_firefox.usage_ping.enabled": False}). Wrapper-level network calls that are not browser launches (CloakBrowser: update checks, licence and GeoIP lookups; Clearcote: licence heartbeat when a key is set, GeoIP lookups, the Widevine CDM download that is on by default for licensed launches) are described on their pages. invisible_playwright leaves Firefox's own telemetry and background services on by design; see its page. | | Prerelease status | Clearcote's browser is v0.1.0-pre.23; Camoufox's current line is v156.0.1-beta.38. Both projects label their builds as non-final. (On 2026-10-13 GitHub no longer flagged Clearcote's pre.23 release as a pre-release, so STATUS.md shows it without the marker; the tag name and the project's own text still call it a pre-release.) |

Star counts appear in STATUS.md for completeness. They measure repository popularity and have no established relationship to detection outcomes.

Capability-to-implementation index

Maps a technical requirement to the tools whose source implements it, per the architecture tables. This is a lookup, not a ranking; where more than one tool qualifies, they are listed in table order.

| Requirement | Tools implementing it | Mechanism | |-------------|----------------------|-----------| | Fingerprint spoofing not observable as JS property-descriptor modification | Camoufox, CloakBrowser, Clearcote, invisible_playwright | Values substituted in C++ before JS can inspect them | | Statistically distributed fingerprint rotation | Camoufox | fpgen-generated profiles with coherence rules (PyPI 0.5.8 and repo HEAD) | | Import and present a specific captured device fingerprint | Clearcote | --fingerprint-profile with a bundled verifier (Camoufox ships opt-in real presets instead, fingerprint_preset=True) | | Unmodified Playwright API surface | Patchright, CloakBrowser, Clearcote | Drop-in driver or SDK | | Runtime.enable tell addressed | Patchright, Clearcote, Camoufox (not applicable — Juggler) | Isolated contexts; C++ patch; non-CDP protocol | | TLS/JA3 impersonation on HTTP requests | Scrapling, Obscura (build-gated) | curl_cffi; wreq | | Mouse-motion model for input | Botasaurus, CloakBrowser, Clearcote, Camoufox, invisible_playwright | Bézier trajectories; C++ humanized input plus SDK paths; recorded-movement replay in Juggler JS; client-side seeded path planner | | CAPTCHA handling in-framework | SeleniumBase, Scrapling (Cloudflare only), Botasaurus (Cloudflare only) | Click-based solving; solve_cloudflare | | HTML parsing without a separate library | Scrapling, Obscura | lxml-based parser; html5ever | | Crawler/spider framework | Scrapling, Botasaurus | Scrapy-like spiders; framework primitives | | Node.js or .NET client | Patchright, CloakBrowser, Clearcote | Published SDKs | | Test-framework integration | SeleniumBase | pytest, unittest | | Rebuildable engine from published sources | Camoufox, Clearcote, Obscura | Patches + build scripts; Rust workspace | | Signed, checksummed release binaries | Clearcote, CloakBrowser | Clearcote: GPG signature + SHA-256. CloakBrowser: the wrapper verifies a pinned Ed25519 signature over SHA256SUMS (the binary itself is closed) | | Resident memory below a browser's | Scrapling (HTTP tier), Obscura | No browser process; own engine — Tier B, vendor figures | | Firefox rather than Chromium engine | Camoufox, invisible_playwright | Firefox 156.0.1 / 151 forks | | WebRTC UDP routed through a SOCKS proxy (implemented, off by default) | invisible_playwright | SOCKS5 UDP ASSOCIATE layer added to Necko (+423 lines), gated by network.proxy.socks_remote_udp (default false), which invisible_core does not set; behind a proxy the core blocks direct UDP and declares a server-reflexive candidate from the exit IP. Clearcote's SDK has a socks5Udp option whose engine side is not in the public tree | | Playwright API shape without a bespoke launcher API | invisible_playwright | Engine carries Juggler; the Python client is a vendored, modified copy |

Requirements with no tool in this set: native macOS Clearcote builds (the SDK runs the Linux Docker image on macOS), WebGL in Obscura (getContext('webgl') returns null), first-party handling of Akamai/DataDome/Kasada/Imperva in any of the ten.

Detection layers and which tools address them

Layer taxonomy with the tools whose source addresses each, per the architecture tables.

Layer 1 — Automation protocol tells
  Runtime.enable / execution contexts .... Patchright, Clearcote
  Non-CDP automation protocol ............ Camoufox (Juggler)
  ChromeDriver binary markers ............ SeleniumBase (cdc_ removal)
  CDP input characteristics .............. CloakBrowser (Tier B), Clearcote
Layer 2 — Browser fingerprinting
  navigator.webdriver .................... all ten
  Canvas / WebGL ......................... Camoufox (WebGL; canvas readback is a random
                                           placeholder only when the claimed OS/GPU is not
                                           the host's), CloakBrowser (Tier B), Clearcote
  Screen / window geometry ............... Camoufox, CloakBrowser (Tier B), Clearcote
  AudioContext ........................... Camoufox, CloakBrowser (Tier B), Clearcote
  Font enumeration ....................... Camoufox, CloakBrowser (Tier B), Clearcote
Layer 3 — Behavioural signals
  Mouse-motion model ..................... Botasaurus, CloakBrowser, Clearcote, Camoufox,
                                           invisible_playwright
  Click and keystroke timing ............. Botasaurus, SeleniumBase, CloakBrowser
  Navigation patterns .................... not addressed by any tool; a property of
                                           calling code, not the driver
Layer 4 — Network
  TLS fingerprint (JA3/JA4) .............. Scrapling (HTTP tier), Obscura (build-gated)
  WebRTC / UDP leakage ................... Camoufox, Clearcote
  IP reputation .......................... not addressed by any tool
  DNS leakage ............................ not addressed by any tool
Layer 5 — Layout and rendering probes
  getBoundingClientRect .................. all real-browser tools; Obscura in
                                           render-enabled builds only
  getComputedStyle ....................... all real-browser tools; Obscura in
                                           render-enabled builds only
  Canvas/WebGL pixel output .............. all real-browser tools; Clearcote can forward
                                           canvas operations to real GPU hardware

Layers 3 (navigation patterns), 4 (IP reputation, DNS), and any account-history signal are outside what a browser-automation library controls. They are properties of infrastructure and calling code.

Variables affecting outcomes

The previous revision published a table of success-rate percentages by protection tier («90%+», «60-80%», «20-40%», «<20%»). No study, target list, sample size, or date was attached to those figures, and they have been removed rather than revised.

What can be stated without a benchmark is which variables the tool layer does and does not control:

| Variable | Controlled by the tools compared here? | |----------|----------------------------------------| | Browser fingerprint surface | Yes — the primary function of Camoufox, CloakBrowser, Clearcote | | Automation protocol tells | Yes — Patchright, Clearcote, SeleniumBase, Camoufox | | Input event characteristics | Partially — five tools implement a motion model | | TLS fingerprint | Only on HTTP-only paths (Scrapling, Obscura) | | IP address reputation | No — supplied by your proxy provider | | Request volume, timing, and path order | No — a property of your calling code | | Account age and session history | No — outside the browser layer | | Target's protection tier and configuration | No — and vendor claims rarely state which tier was tested | | Durability over time | No — detection is adversarial; a passing result is scoped to that target on that date |

Consequence for evaluating any claim, in this repository or elsewhere: a stated pass rate without the IP type, target list, protection tier, trial count, and date is underspecified, because the unstated variables include ones the tool does not control.

Detection test sites

Consistency auditors and vendor demos. Passing them is not evidence about production WAF deployments.

| Test | Checks | URL | |------|--------|-----| | Sannysoft | Automation tells | bot.sannysoft.com | | BrowserScan | Fingerprint surface | browserscan.net | | CreepJS | Fingerprint coherence and contradiction detection | abrahamjuliot.github.io/creepjs | | Fingerprint.com | Vendor bot-detection demo | fingerprint.com | | Pixelscan | Leak and consistency detection | pixelscan.net |

Reproducing this report

Version and health data:

python scripts/verify.py --write

Source verification. The ten projects are unaudited third-party code, including patched browser binaries; they are cloned into a container and never onto a workstation:

./scripts/sandbox.sh up        # clone into a Docker volume, then disconnect the network
./scripts/sandbox.sh sh        # offline shell for reading source
./scripts/sandbox.sh nuke      # destroy container and volume

The container runs with --cap-drop=ALL, --security-opt=no-new-privileges, a read-only root filesystem, and no network after cloning completes. No dependency installation, build, test run, or binary execution is performed at any point in producing this report; Tier A means the code was read, not run. Details: METHODOLOGY.md § Source handling.

Contributing

To change a claim, open an issue quoting it and supplying one of:

• a file path and commit that contradicts it (Tier A);
• vendor documentation that supersedes it (Tier B);
• a reproducible test stating target, IP type, protection tier, trial count, and dates

To propose a tool: repository URL, modification method, and the claims it publishes.

Scope and legal

This document describes publicly available software for security research, academic work, authorised testing, and retrieval of one's own data. Respect robots.txt, rate limits, and applicable terms of service. Laws governing automated access vary by jurisdiction; this is not legal advice.