How the real-browser engine works

The exact fingerprint stack every Browsa browser emits. Published so you can audit it, not just trust the marketing copy.

The stack at a glance

Layer	What we patch	Where
navigator	UA, platform, languages, vendor, webdriver, hardware_concurrency, deviceMemory	Chromium C++ patch + V8 binding override
screen	width, height, availHeight, colorDepth, pixelDepth, devicePixelRatio	Chromium C++ patch
WebGL	UNMASKED_VENDOR, UNMASKED_RENDERER, getShaderPrecisionFormat, WEBGL_debug_renderer_info, extensions	Chromium C++ patch with per-OS values (Apple Inc. / ANGLE-Apple-M2 for macOS, Google Inc. / Intel UHD for Windows)
Canvas	Per-session deterministic noise on toDataURL / getImageData (canvas-session-ID plumbed end-to-end per profile)	Chromium C++ patch (Skia path)
Audio	AudioBuffer noise injection, hardware-coherent baseLatency derived from sample_rate / buffer_size	Chromium C++ patch (Blink AudioContext)
Fonts	Real OS font list (not the "Docker SwiftShader stub" that ordinary automation ships)	Phantom Docker image includes the real macOS/Windows font set
CDP signal	navigator.webdriver = false, all cdc_* properties stripped, Runtime.evaluate non-detection	Chromium C++ patch + STEALTH_MODE flags
TLS / JA3	Client Hello fingerprint matches real Chrome 147 — utls MITM proxy emits the same SNI extension order, GREASE values, cipher list	Phase 4 sidecar (ja3-proxy / phantom-forwarder)
HTTP/2 settings	SETTINGS frame order + values match Chrome 147	Phase 4 sidecar
Timezone / locale	Derived from proxy.country at burner boot — primary_language + accept_languages + Intl.DateTimeFormat all coherent	cloud-runner boot-time env injection

Per-OS values

The exact strings we emit per os_type. GET /v1/agents/{id}/fingerprint returns this shape for any provisioned agent.

OS_TYPE_MACOS (default)

navigator.userAgent	Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/147.0.7727.25 Safari/537.36
navigator.platform	MacIntel
WebGL UNMASKED_VENDOR	Apple Inc.
WebGL UNMASKED_RENDERER	ANGLE (Apple, ANGLE Metal Renderer: Apple M2, Unspecified Version)
screen.colorDepth	30 (matches real M-series Macs)
screen.pixelDepth	30
hardware_concurrency	8
deviceMemory	8

OS_TYPE_WINDOWS

navigator.userAgent	Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/147.0.7727.25 Safari/537.36
navigator.platform	Win32
WebGL UNMASKED_VENDOR	Google Inc. (Intel)
WebGL UNMASKED_RENDERER	ANGLE (Intel, Intel(R) UHD Graphics 630 Direct3D11 vs_5_0 ps_5_0, D3D11)
screen.colorDepth	24
screen.pixelDepth	24
hardware_concurrency	8
deviceMemory	8

Realism-harness results

We run every Phantom build through an automated realism harness against the major realism-check sites. Current pass rate on a fresh burner (no warmup, US proxy):

Detector	Verdict	Notes
antoinevastel.com/bots	PASS	Full clean — reads as an authentic browser, no headless signal
fingerprint.com BotD (open source)	PASS	bot: false with STEALTH_MODE
fingerprint.com Pro demo	PASS	Suspect Score 29 (≤30 = not flagged)
browserleaks.com	PASS	Clean across UA, fonts, WebGL, audio
creepjs (abrahamjuliot)	SUSPICIOUS	Trust ~38% — improvement on persistent identities with warmup recipe
iphey.com BROWSER	SUSPICIOUS	Missing-API signals (navigator.share, contentIndex, contacts) — fix in next Phantom build

Trust Score reflects this signal: burner default ~870-890; persistent identities with warmup land in the 900-940 range. See trust_score docs.

Power-user overrides

Most customers should let the platform pick a coherent combination (this is what Trust Score is calibrated against). When you need finer control, six override knobs are exposed at agent-creation time. Empty = platform default; invalid values silently degrade to the default rather than refusing the provision.

Override	Accepted values	Default
screen_resolution	1366x768, 1440x900, 1536x864, 1600x900, 1920x1080 (default), 2560x1440, 3840x2160, mobile-iphone (390x844), mobile-android-pixel (412x915), or custom WxH (320≤W,H≤7680)	1920x1080
webrtc_mode	masked · real · disabled	masked
geolocation_mode	prompt · allow · block	prompt
geolocation_data	"lat,lng" string (lat∈[-90,90], lng∈[-180,180]) — used when geolocation_mode="allow"	derived from country
media_devices_mode	masked · real	masked
port_scan_protection	masked · real	masked

Set them on the agent at create time:

curl https://browsa.io/v1/agents \
  -H "X-API-Key: agt_live_..." \
  -d '{
    "mode": "burner",
    "os_type": "OS_TYPE_MACOS",
    "country": "de",
    "screen_resolution": "2560x1440",
    "geolocation_mode": "allow",
    "geolocation_data": "52.5200,13.4050"
  }'

Coherence matters. The platform doesn't refuse incoherent combinations (some advanced use cases want a mobile resolution on a macOS UA), but it surfaces them in the warnings array on GET /v1/agents/{id}/fingerprint:

{
  "screen": { "width": 390, "height": 844, ... },
  "warnings": [
    "screen_resolution looks mobile-shaped but os_type is desktop — realism checks cross-check; trust_score may drop",
    "media_devices_mode=real returns the host's device list — in Phantom Docker this is 0 devices, which is itself a realism tell"
  ],
  ...
}

Verify it yourself

Four ways to audit the fingerprint without taking our word for it:

1. GET /v1/agents/{id}/fingerprint — the wire-level preview for any provisioned agent.
2. GET /v1/agents/{id}/trust — the per-factor trust-score breakdown.
3. GET /v1/agents/{id}/detection-signal — the most-recent realism-harness verdict for your identity's fingerprint shape (sites_tested + sites_passed + per-site verdicts). The empirical data behind the score number.
4. Connect to the agent via CDP (c.connect(agent_id) in the SDK) and run any realism-check URL through it — we publish that we pass; you can confirm it.

from browsa import Client
c = Client(api_key="agt_live_...")

agent = c.create_agent(country="us", os_type="OS_TYPE_MACOS")
print(c.fingerprint(agent["id"]))  # full wire spec
print(c.trust_breakdown(agent["id"]))  # why the score is what it is
print(c.connect(agent["id"]))  # paste-ready Playwright snippet

Start free — 300 credits