shardx (Python)

Self-contained Python SDK for the ShardX anti-detect browser by the ProxyShard team.

This package does not depend on the desktop launcher. On first use it downloads the patched Chromium 152 engine, Widevine CDM, and the 220-profile fingerprint library from our CDN into a local cache, then launches isolated browser sessions on demand.

Driven by patchright (stealth-patched Playwright) — sdk.session() returns a ready-to-use Browser instance, no manual connect_over_cdp plumbing.

Install

pip install shardx

Supported hosts: macOS arm64, Windows x64, Linux x64.

Linux system dependencies

The bundled Chromium engine needs unzip + the standard set of shared libraries any Chromium fork links against. On a fresh Debian / Ubuntu:

sudo apt install -y \
  unzip ca-certificates fonts-liberation \
  libnss3 libnspr4 libatk1.0-0 libatk-bridge2.0-0 libcups2 \
  libxkbcommon0 libxcomposite1 libxdamage1 libxfixes3 libxrandr2 \
  libgbm1 libpango-1.0-0 libcairo2 libasound2 libxshmfence1

When launching as root or inside Docker, pass --no-sandbox and --disable-dev-shm-usage via extra_args=:

await sdk.session(..., extra_args=["--no-sandbox", "--disable-dev-shm-usage"])

Quick start

import asyncio
from shardx import ShardX

async def main():
    sdk = ShardX()
    # Engine + Widevine + fingerprint library auto-download from CDN on
    # the first `session`/`launch`/`list_profiles` call (~170 MB once,
    # etag-cached afterward).  No separate install step.

    # Create a persistent profile from a library template (or create_profile()
    # for a random one). Library templates aren't launched directly — this
    # freezes an enriched copy under a unique id you can return to. Do it once.
    profile = sdk.create_profile("win-rtx4060")

    # Launch + drive in one call. Yields a patchright `Browser`.
    async with sdk.session(profile, proxy="socks5://user:pass@host:port") as browser:
        ctx = browser.contexts[0]
        page = await ctx.new_page()
        await page.goto("https://browserleaks.com/quic")
        print(await page.title())

        # Inspect what the SDK resolved before launch:
        sess = browser._shardx
        print(sess.geo)                   # GeoInfo(...) from ip-api / ipapi.co
        print(sess.proxy_udp_ms,          # UDP RTT in ms or None
              sess.quic_enabled,          # bool
              sess.webrtc_mode)           # "auto" | "tcp_only" | "block"
    # browser + udd shut down cleanly on exit

asyncio.run(main())

Random profile

# create_profile() with no id freezes a random library template (filter the
# pool with platform="Windows"). It already randomises hardware_concurrency /
# device_memory / platform_version once, at creation.
profile = sdk.create_profile(platform="Windows")
async with sdk.session(profile) as browser:
    page = await browser.contexts[0].new_page()
    ...

Discover bundled profiles

print(sdk.list_profiles()[:5])
# ['linux-gt1030', 'linux-gtx1050', 'mac-m1-air13', 'mac-m1-imac24', 'mac-m1-max-mbp14']

print(sdk.list_profiles(platform="Windows")[:5])

profile = sdk.random_profile(platform="macOS")
print(profile.id, profile.config["webgl"]["renderer"])

Validate a proxy before binding

print(sdk.check_proxy("socks5://user:pass@host:port"))
# {
#   'udp_ms': 142.3,
#   'geo': GeoInfo(country_code='DE', timezone='Europe/Berlin', ...),
#   'would_enable_quic': True,
#   'would_set_webrtc': 'auto',
# }

Persistent profiles

list_profiles() / random_profile() hand back library templates — launch one and it re-reads the same template every time. For a profile you can return to (same fingerprint and cookies/cache) or delete, create a saved profile: it freezes a template (or a random one) with randomized hardware/platform_version under a fresh unique id in its own folder <cache>/profiles/<id>/, exactly like the desktop launcher's "create profile".

sdk = ShardX()

# Create once (random template, or pass a library id like "win-rtx4060").
profile = sdk.create_profile()
print(profile.id)

print(sdk.list_saved_profiles())          # ['<id>', ...]

# Launch it. randomize stays False → the frozen fingerprint is stable;
# cookies/cache persist in the profile's folder across runs.
async with sdk.session(profile) as browser:
    ...

# Later — even a different process — reopen by id: same fingerprint + state.
profile = sdk.open_profile("<id>")

# Remove the profile and all its state when you're done.
sdk.delete_profile("<id>")

Saved profiles never touch the bundled S3 library: templates stay read-only in <cache>/fingerprints/*.json, saved profiles live in <cache>/profiles/<id>/ (holding profile.json + the browser's user-data-dir).

Human input and finger gestures

Motion is the engine's own input domain: pointer trajectories that obey Fitts's law, key-by-key typing with log-normal gaps, and the gestures a cursor cannot make. It runs inside the browser process — nothing is injected into the page — and it does not appear in /json/protocol, so a page cannot discover it.

from shardx import ShardX, Motion

async with sdk.session(profile) as browser:
    page = await browser.new_page()
    # Motion acts on a page's widget, so it attaches to a page.
    motion = await Motion.attach(page)

    # desktop profile
    await motion.create_pointer(20, 20)
    await motion.glide_to(640, 360, target_width=220)   # the field's real width
    await motion.tap()
    await motion.enter_text("hello there")
    await motion.destroy_pointer()

    # phone profile
    await motion.touch_swipe(200, 600, 200, 200, flick=True)   # scroll and fling
    await motion.touch_long_press(200, 300)                    # context menu
    await motion.pinch(200, 400, 2)                            # zoom in
    o = await motion.set_orientation(90)
    # o → {'angle': 90, 'type': 'landscape-primary', 'screenWidth': 844, ...}

A profile that claims a touchscreen has no cursor, and the core enforces it both ways: the pointer commands are refused on such a profile, and the finger commands are refused on every other. There is no touchScroll and no fling command — gestures are built in the browser out of the touch stream, so a real swipe produces the scroll and the fling with the velocity the browser's own tracker fitted. touch_drag differs from touch_swipe in the wait before the travel: that wait is when a page's drag-and-drop starts, and a stroke that begins earlier is a scroll.

set_orientation is physical and takes time — the sensors move first and the picture commits at the end, the order a handset produces. It returns once the new angle has reached the page, so screen.width read straight after is already the turned one.

Anti-fingerprint noise

Per-vector noise (canvas / WebGL / audio / DOMRect / sensors / fonts) is off by default. set_noise(...) is declarative — exactly the vectors you list are enabled (with soft defaults), every other one is turned off:

profile.set_noise("canvas", "audio", "webgl")   # only these three on
profile.set_noise("canvas")                       # audio + webgl now off again
profile.set_noise()                               # all off
sdk.save_profile(profile)                          # persist the choice

Seeds are derived per-profile at launch — stable across runs, unique per profile — so two profiles with the same vectors enabled still produce different canvas/audio/WebGL fingerprints. Soft defaults: WebGL intensity 0.0005, DOMRect max_offset 1.

Pre-launch checks

Every call to sdk.session() / sdk.launch() runs the same pre-spawn pipeline the desktop launcher uses:

1. resolve_auto_fields — if the profile has "auto" sentinels for timezone, navigator.language, or geolocation.mode, the SDK makes a live geo lookup through the bound proxy (ip-api.com by default). It then writes concrete values: timezone (from the API, never a static table), accept_language chain, languages, icu_locale (always overwritten so Intl.* matches navigator.language), and lat/lng. Proxy-via failure → direct geo → host $LANG / $TZ as last-resort fallback. The chosen geo is surfaced on session.geo.
2. apply_screen_strategy — see below.
3. probe_udp — SOCKS5 UDP_ASSOCIATE round-trip. If it fails, QUIC is force-disabled and WebRTC switches to tcp_only automatically.

Screen strategy

screen_mode kw to session() / launch():

• "profile" — keep whatever the fingerprint claims.
• "cap_to_host" — macOS default. If the host monitor is smaller than the FP screen, scale screen.* + window.* down proportionally; otherwise no-op.
• "use_host" — Windows/Linux default. Overwrite screen.* with the real monitor (minus a 40 px Windows taskbar) and recompute window.outer_* / window.inner_* accordingly.

Default mode is picked from navigator.platform. Override per launch:

async with sdk.session(profile, screen_mode="profile") as browser:
    ...

Host-aware hardware randomisation

randomize=True re-picks hardware_concurrency, device_memory, and platform_version before the launch — using the same logic as the desktop launcher (randomize_hardware in lib.rs):

• macOS profiles use the curated MAC_HW_CONFIGS table by id.
• Windows / Linux profiles bracket the host's logical CPU count within [host − 4, host + 2] from the real x86 set [4, 6, 8, 12, 16, 20, 24, 28, 32]; device_memory is floored by core count (≥ 12 → 16, else 8) and capped by host_ram_bucket_gb() (8 / 16 / 32 GiB bucketed from sysctl hw.memsize / /proc/meminfo / Get-CimInstance Win32_ComputerSystem).

So a profile launched on an 8-core / 16 GB laptop will never claim 32 cores / 128 GB of RAM — keeps fingerprints internally consistent with real-world hardware.

Override fingerprint fields

profile = sdk.library.load("win-rtx4060").with_override(
    name="my-account",
    timezone="Europe/Berlin",
    navigator={"language": "de-DE"},
)
async with sdk.session(profile, proxy="socks5://...") as browser:
    ...

Use your own fingerprint JSON

from shardx import Profile

profile = Profile.from_file("/path/to/my-custom.json")
async with sdk.session(profile) as browser:
    ...

WebRTC policy

async with sdk.session(
    profile,
    proxy="socks5://...",
    webrtc="tcp_only",                # or "block" | "auto" (default)
    webrtc_public_ip="203.0.113.42",  # advertised in ICE candidates
) as browser:
    ...

Advanced: raw launch without patchright

If you'd rather drive the browser with a different CDP client (raw pychrome, pyppeteer, your own WebSocket), skip session() and use launch() directly:

profile = sdk.create_profile("win-rtx4060")
sess = sdk.launch(profile, proxy="socks5://...", cdp=True)
print(sess.cdp_url)        # ws://127.0.0.1:54113/devtools/browser/c0a3…
# … drive it yourself …
sess.stop()

launch() does the same pre-launch pipeline (auto-resolve, screen strategy, UDP probe, hw randomisation) and returns a BrowserSession with cdp_url, geo, proxy_udp_ms, quic_enabled, webrtc_mode, user_data_dir, and stop().

Cache layout

~/Library/Application Support/shardx-sdk/    (mac)
%LOCALAPPDATA%\shardx-sdk\                   (win)
~/.cache/shardx-sdk/                         (linux)
├── manifest.json             ← etag cache for browser/widevine/fingerprints
├── ShardX-Mac-arm64/         ← extracted engine
│   └── ShardX.app/…
├── fingerprints/             ← 220 bundled .json profiles
│   ├── win-rtx4060.json
│   └── …
└── profiles/                 ← saved profiles (create_profile) + their state
    └── <profile-id>/
        ├── profile.json      ← the frozen fingerprint config
        └── …                 ← user-data-dir: cookies, IndexedDB, cache

Override the cache root:

sdk = ShardX(cache_dir="/data/shardx")

Update the runtime

The SDK auto-checks remote etags on the first session/launch/list_profiles call of each process and re-downloads anything that changed. To force a re-download mid-process (e.g. CI scenarios):

sdk.runtime.install(force=True)

License

MIT (this SDK). The Chromium-fork engine binary downloaded at runtime is a closed-source product — see the main repo for engine licensing.