What it is

Pydoll is a Python async browser automation library built directly on the Chrome DevTools Protocol (CDP). It was created by Thalison Fernandes under the autoscrape-labs organization. Unlike Playwright or Selenium, Pydoll has no WebDriver layer: it communicates with Chrome directly via CDP. It is async-first by design.

How it works

Pydoll opens and controls Chrome through raw CDP commands. The async architecture allows concurrent tab management, making it theoretically suitable for multi-tab scraping workloads without the overhead of separate browser processes.

The library provides browser contexts for session isolation and exposes human-like interaction primitives: click offset randomization, keystroke delays, and smooth scroll behavior.

A notable feature is expect_and_bypass_cloudflare_captcha(), a method that handles Cloudflare Turnstile challenges. The library’s documentation also references future features including Bezier curve mouse paths, physics-based scrolling, and cognitive delays between actions, positioned as a behavioral analysis bypass layer.

TWSC experience

We tested Pydoll in 2026 as part of the Cloudflare bypass benchmark. Results were poor across both targets:

  • Indeed: 0% success. Pages loaded incompletely and the tool failed silently without raising exceptions.
  • Harrods: 1% success. CDP connections crashed repeatedly, with Chrome consuming 98.9% CPU before dying.

We also found significant documentation accuracy issues. The Browser class referenced in multiple code examples in the official documentation does not exist. tab.html() does not work as documented. tab.current_url is an attribute, not a method, contrary to the documentation showing it called as tab.current_url().

The future features listed in the documentation (Bezier curves, physics scrolling, cognitive delays) had not shipped at the time of testing.

Known limitations

  • Unreliable under real anti-bot conditions. Both test targets produced near-zero success rates.
  • Chrome resource consumption is severe. The 98.9% CPU figure makes concurrent deployment impractical.
  • Documentation does not accurately reflect the current API. Code copied from the docs will not run without debugging.
  • Silent failure mode on incomplete pages makes error detection harder than with exception-throwing tools.
  • Promised behavioral features were not available at time of testing.

Sources


The Web Scraping Wiki is compiled from the tests and research published on The Web Scraping Club, the newsletter for web scraping professionals. Subscribe to get the next article in your inbox.