Skip to main content

sleight

Drive any CDP browser like a human. Bezier trajectories with real hand tremor, typing rhythm modelled on keystroke-dynamics research, and exclusive leasing for browser instance pools.

Python ≥ 3.11 · one runtime dependency (websocket-client) · MIT

中文 README · 📖 Wiki — full documentation

pip install sleight                     # core: one dependency
pip install "sleight[xpath]"            # + lxml, for parse(xpath=True)
pip install "sleight[ui]"               # + fastapi/uvicorn, for `sleight ui`
pip install "sleight[redis]"            # + redis, cross-process leasing

sleight browser install fingerprint-chromium   # optional: an anti-detect kernel

30 seconds

from sleight import connect, Text

with connect("http://127.0.0.1:9222") as s:      # opens its own tab, closes it on exit
    s.open("https://example.com", wait=Text("Example Domain"))
    print(s.title(), len(s.content()))

With a browser pool that has real profiles behind it:

from sleight.providers import CloakBrowserManager

mgr = CloakBrowserManager("http://127.0.0.1:19000", token="…")

with mgr.lease() as inst:                    # exclusive lease, released on exit
    with inst.session(human=True) as s:      # every action gets a human trajectory
        s.open("https://example.com")
        s.click("#login")
        s.type("#email", "user@example.com")
        s.click("#submit", human=False)      # …except this one, speed matters here

See what the page actually loaded — the library gives you structured data, you decide what to print:

with s.capture_resources(types={"Script", "Stylesheet"}) as capture:
    s.open(url, wait=Load())
    s.pump_events(10)          # the async batch that arrives after `load`

for r in capture.snapshot():
    print(r.resource_type, r.status, r.url)

Target one specific profile — by id, by name, or by tag. A name that matches nothing fails immediately with the visible names listed, rather than blocking until timeout:

with mgr.lease(instance_id="5edcc28a-…") as inst:              ...
with mgr.lease(name="Win-US-02") as inst:                      ...
with mgr.lease(where=lambda i: "us" in i.tags) as inst:        ...

handles = pool.lease_many(4, names=NAMES, timeout=60)   # rolls back on partial failure

Drag a slider — the buttons mask stays down for the whole haul, the trajectory overshoots and comes back, and there is a pause before the release, because releasing the instant you arrive is the most reliable machine tell there is:

s.drag("#captcha-knob", by=(212, 0), human=CAREFUL)
s.drag_and_drop("#card", "#done-column")     # HTML5 native drag, or a JS one — both

Rotate the exit IP. The tunnel hands out addresses per TCP connection and Chrome reuses keep-alive sockets, so a whole run pins to one IP. A fresh browser context is the only thing that reliably breaks that — clearing cache, unique query strings, and emulateNetworkConditions all do nothing (why):

with inst.context() as ctx, ctx.session() as s:   # own socket pool → new exit
    print(ctx.exit_ip())
    s.open(url)

Stop paying for bytes you throw away, and shed the tracking cookie afterwards:

with s.block(types=["Image", "Media", "Font"]) as blocked:
    s.open(url)
print(blocked.by_type)                       # {'Image': 34, 'Font': 6}

report = s.clear_site_data("https://example.com")
print(report.cookies)                        # ('datadome',) — what actually went away

Three providers' worth of instances, one logical pool:

from sleight import Pool
from sleight.providers import CloakBrowserManager, Plain

pool = Pool([
    CloakBrowserManager("http://10.0.0.1:9000", token=T1, name="hk"),
    CloakBrowserManager("http://10.0.0.2:9000", token=T2, name="sg"),
    Plain("http://127.0.0.1:9222", name="local"),
])

with pool.lease(where=lambda i: "us" in i.tags) as inst:
    ...

Driving it from an LLM

Compress the page into something a model can read, and get a stable ref back for every interactable element. Clicking a ref goes through the same human input chain — real isTrusted events, both hit tests, everything:

snap = s.snapshot()
print(snap.text())
# RootWebArea "Checkout"
#   heading "Your order"
#   textbox "Card number" [e1]
#   button "Pay now" [e2]

s.type(snap.ref("e1"), "4242 4242 4242 4242")
s.click(snap.ref("e2"))

Refs are keyed by (loaderId, backendNodeId): the same node keeps the same ref across snapshots, and every ref dies the moment the page navigates (StaleRef) rather than silently pointing at whatever now sits in that slot. Same-process iframes are merged in, so an element inside an iframe gets a ref like any other.

Pull the article and the usual fields out of the rendered DOM — no lxml, no Node:

doc = s.extract_document()
doc.title, doc.byline, doc.text[:80], doc.json_ld, doc.low_quality

Reading a lot of elements? Take one static snapshot and query it in memory instead of paying a CDP round-trip per element (~18× faster on a 50-row table):

dom = s.parse()                                   # one fetch, then pure Python
names = [e.text for e in dom.query_all("tr.row td.name")]
deep  = s.parse(pierce_shadow=True)               # open shadow DOM inlined too

Launch a local browser yourself — any Chromium build, including an anti-detect one. Same seed, same fingerprint, every run:

from sleight import launch

with launch("fingerprint-chromium", fingerprint=42) as s:
    s.open("https://example.com")

And the whole thing is an MCP server, so a model can drive it directly:

SLEIGHT_CDP_URL=http://127.0.0.1:9222 sleight-mcp

It speaks JSON-RPC over stdio with no extra dependencies and exposes four tools — browser_session / browser_observe (snapshot + find) / browser_act (by ref or by coordinate, for canvas and icon-only UIs) / browser_extract. Raw CDP and eval stay hidden unless you opt in.

Getting a fleet to drive

The hard part of running CloakBrowser is not the driving — it is the deployment, the extension rollout, and the "why does this profile behave differently" archaeology. So the same package ships a CLI for it. Local docker or a remote host over SSH is the same code path, only the runner differs:

sleight hosts add hk-01 --ssh deploy@10.0.0.12 --dir /srv/cloakbrowser-manager --sudo
sleight deploy --host hk-01
sleight deployments add hk-01 second --dir /srv/cbm-2 --port 9001   # same box, second manager
sleight ext push ./plugins/bypass-paywalls --host hk-01   # MV3 check + permissions
sleight ext apply --host hk-01                            # every profile, then restart
sleight ext verify --host hk-01                           # did the browser really load it
sleight ui                                                # the same, in a browser

Hosts, the managers on each of them, and a deploy/backup/upgrade audit trail live in a local SQLite database (~/.sleight/sleight.db) that the CLI and the web UI share.

sleight ui walks you through it: connect a host (with a real connection test before anything is saved), pick a sizing template, preflight, deploy. Every option carries a one-line explanation of what breaks if you get it wrong plus a recommended value — defined once on the backend, rendered by both the CLI (sleight templates) and the UI.

Deploys are idempotent, --dry-run prints the exact bytes it would write, and the things you must not do are refused rather than documented: no latest tag, no down -v, no silently rotating an in-use AUTH_TOKEN, no second manager on the same /data. The engine is stdlib-only (SSH is the system ssh binary); only sleight ui needs pip install "sleight[ui]".

Why this exists

Fingerprint-level anti-detection is a solved problem — CloakBrowser patches Chromium at the source level, Camoufox patches Firefox. They fix what the browser looks like. Nothing fixes how it moves.

  • Playwright and Puppeteer teleport the mouse. mouse.move(steps=N) interpolates a straight line at constant speed — zero jitter, zero acceleration. That is itself a signature.
  • The browser will not fill in the trajectory for you. Even with a humanize feature enabled browser-side, an external CDP client produces zero mousemove events between press and release. Measured, not assumed.
  • The good trajectory work lives in JavaScript (ghost-cursor). Python ports are thinly maintained.
  • Crawlee for Python's BrowserPool does not support remote browsers.

sleight fills exactly that gap: Python + remote CDP + human behaviour + instance leasing.

Relationship to Playwright

Not a replacement — a complement. sleight is a driver layer, not a framework. It deliberately does not do downloads, video, tracing, or a full locator DSL. When you need those, use Playwright.

The interesting part is that you can use both: sleight's human module is sans-io — it emits (method, params, sleep_after) tuples and never touches a socket — so it drives a Playwright CDPSession just as happily as sleight's own transport.

What makes the motion credible

sleight typical automation
Path shape cubic Bezier, control points offset to one side straight line
Micro-motion WindMouse wind term (correlated tremor) none, or white noise
Point count Fitts's law — far small targets take longer fixed steps=N
Landing truncated Gaussian inside the box dead centre
Coordinates integers floats used as "jitter"
Overshoot past the target then back, distance-scaled exact arrival
Typing per-character events, interval by digraph class one insertText
Scrolling repeated small mouseWheel deltas one scrollTo
Dragging buttons mask held the whole way, slider-grade overshoot, pause before release teleport, or release on arrival

Parameters are not invented. They come from the WindMouse physical model, ghost-cursor's Fitts-law point budgeting, and published keystroke-dynamics measurements (alternating-hand digraphs average 114 ms, same-hand-different-finger 131 ms, same-finger slowest and most variable).

Scope

Does: navigation, reload and history · typed wait conditions · rendered-DOM reads · CSS queries · human mouse / keyboard / wheel / drag · element screenshots · forms (select_option, upload_file) · isolated browser contexts for exit-IP rotation · origin-scoped site-data clearing · request blocking via the Fetch domain · exit_ip() · structured network-resource capture · instance discovery across providers · cooperative exclusive leasing with TTL renewal (in-memory, or Redis-backed across processes) · idempotent recovery · deploying and operating CloakBrowser Manager over local docker or SSH, extensions included.

Also does (the LLM-facing layer, see below): iframe / OOPIF / Shadow DOM piercing · accessibility snapshots with stable refs · main-content and metadata extraction · a four-tool agent gateway and an MCP server · launching a local browser binary.

Does not: scheduling and queues · fingerprint spoofing (that is the browser's job — sleight drives one, see LocalLauncher) · strict fencing · WebDriver BiDi · Firefox.

The deploy layer lives in its own subpackage and is never imported by import sleight, so the driver stays a one-dependency library.

Known limits

Measured, not assumed. Each of these cost someone a day to find out:

No extensions inside a browser context Target.createBrowserContext makes an off-the-record context, and Chrome does not enable extensions there. Same profile, same URL: chrome-extension://<id>/… opens in the default context and returns ERR_BLOCKED_BY_CLIENT in a fresh one. So rotating the exit IP and using a plugin are mutually exclusive — if your run depends on one, rotate by leasing different profiles with different upstream proxies instead.
reload() on a redirecting URL can return early One redirect is two document commits, and the intermediate one may fire its own DOMContentLoaded. Measured 3/8 on http://https://, 0/8 without the redirect. Not specific to sleight. Wait on something page-specific (Selector, Text) when it matters.
block() only bites while sleight is talking to the browser Paused requests need the event pump, which runs inside open / wait / pump_events / every call. A plain time.sleep() stalls them.
Transport belongs to the thread that created it Enforced, not documented-and-hoped: cross-thread use raises. Lease one instance per thread. Pool and the lease table are shared on purpose.
set_viewport() does not change screen.*, and clear_viewport() may not resize anything The override is render-layer; screen dimensions are a profile fingerprint field fixed at launch. Clearing the override only guarantees no override — measured on Chromium 146 + Xvnc, the window bounced back on half the attempts and stayed at the overridden size on the other half. Set the size you want; do not rely on restoring.

Roadmap

Ordered by what actually blocks work, not by size.

  • iframe / frame supportshipped. s.frames(), s.frame(sel), s.frame_element(iframe, sel), and s.snapshot() merges child frames — both same-process and cross-origin OOPIFs (each via its own CDP session, with the parent-page offset applied) — so elements inside any iframe get refs you can click directly. The DataDome-style slider inside an iframe is reachable now.
  • Context vs. lightweight-instance resource numbers — memory, CPU, and time-to-ready for instance with proxy+plugin / bare instance / N contexts in one instance. Nobody should redesign their concurrency around contexts without this table, so the API stays an opt-in dimension until the numbers exist.
  • launch_args_effectiveget_profile() returns the configured launch args; the effective command line lives on the Manager side. Diagnosing proxy problems currently means reading chrome://version.

Status

0.x — alpha, the API will move. Every release documents its breaking changes in its git tag. Releases are published from that tag by .github/workflows/publish.yml via PyPI Trusted Publishing — no token is stored in this repository.

License

MIT

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

sleight-0.5.0.tar.gz (394.2 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

sleight-0.5.0-py3-none-any.whl (290.6 kB view details)

Uploaded Python 3

File details

Details for the file sleight-0.5.0.tar.gz.

File metadata

  • Download URL: sleight-0.5.0.tar.gz
  • Upload date:
  • Size: 394.2 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for sleight-0.5.0.tar.gz
Algorithm Hash digest
SHA256 8d827b297c4c5457644a53c74d1b663ddf4c13f9f9a8ad41754cf74d930db3b9
MD5 4c2c3e46a4413e57dc787348f1cfee68
BLAKE2b-256 099df6598db3233687468adf362a5e90f9ef01179d57d11db867dd17a4ff2167

See more details on using hashes here.

Provenance

The following attestation bundles were made for sleight-0.5.0.tar.gz:

Publisher: publish.yml on yuanqimanong/sleight

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file sleight-0.5.0-py3-none-any.whl.

File metadata

  • Download URL: sleight-0.5.0-py3-none-any.whl
  • Upload date:
  • Size: 290.6 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for sleight-0.5.0-py3-none-any.whl
Algorithm Hash digest
SHA256 9eeb899f11d9b53c72c93c604bccc37d0bd65d31495a38088b15da912f97cc12
MD5 5bb2f7dc457c9ed8e5236f2a80b7d755
BLAKE2b-256 12f050177bf76f30c707699ac534a15271bdbd4b8bc3e25d3692296ba85da733

See more details on using hashes here.

Provenance

The following attestation bundles were made for sleight-0.5.0-py3-none-any.whl:

Publisher: publish.yml on yuanqimanong/sleight

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Release history Release notifications | RSS feed

This release

0.5.0 This release

2 files

0.4.0

2 files

0.3.0

2 files

0.2.0

2 files

0.1.4

2 files

0.1.3

2 files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page