Benchmark · Measured
Agent-browser compatibility (Lexbench)
Webfold passes all 1,928 tasks of Lexbench. Lexbench is the independent benchmark for headless browsers as agent runtimes. Chrome passes 1,926 in the published run. This is our own run of the suite. It does not cover every website.
Lexbench is the independent test for headless browsers as agent runtimes. 1,928 tasks across 13 drivers. The web platform is covered too. Webfold passes all of them.
What question does this benchmark ask?
Can Playwright, Puppeteer or Selenium drive Webfold's engine as they drive Chrome? Ten more tools are tested. Does the page behave?
How was it run?
- Lexbench-Headless-Browser is published by Lexmount, the makers of the Moli engine, under Apache 2.0. It is not our benchmark. We vendored it unchanged and ran it against Webfold.
- 1,928 tasks in 18 subsets. Layer 1 (1,740 tasks) is protocol and driver compatibility: raw CDP plus 13 pinned drivers in five language ecosystems, including Playwright, Puppeteer, Selenium, chromedp, chromiumoxide, rod, ferrum, pydoll, Stagehand and agent-browser. Layer 2 (188 tasks) is web-platform semantics: DOM, storage, network, workers and CSSOM, judged by what the page finally does, not by a protocol echo.
- Webfold: every attempt passed, in three full runs on the final binary. The other engines' figures are from the benchmark's own published four-engine run of 12 August 2026 (three attempts per task; a task passes only if all three pass).
What are the results?
| Check | Webfold | Reference |
|---|---|---|
| Lexbench, all 1,928 tasks | 1,928 passed | Chrome 1,926 (99.90%), Moli 1,556 (80.71%), Lightpanda 845 (43.83%), Obscura 762 (39.52%), published run |
| Layer 1: protocol and 13 drivers (1,740 tasks) | all passed | includes Playwright, Puppeteer, Selenium, chromedp, rod, Stagehand |
| Layer 2: web-platform semantics (188 tasks) | all passed | DOM, storage, network, workers, CSSOM |
| Ordinary page JavaScript over raw CDP (68 expressions) | 67 match Chrome | the one difference: getComputedStyle(document.body).display |
| Kernel's CDP endpoint contract (14 checks) | 14 of 14 | Chrome 14 of 14 |
| Login with password, wrong code, then TOTP, driven by unmodified Puppeteer (8 checks) | 8 of 8 | |
| WebSocket and EventSource under injected network faults (28 checks) | 28 of 28 | |
| Live accessibility audit on seeded defects (8 defects) | 8 of 8 | Chrome 8 of 8 |
Webfold passes every task. In the benchmark's published run Chrome passes 1,926 and the other lightweight engines between 40% and 81%.
The 188 web-platform tasks are the answer to whether it handles pages that run JavaScript: they check what the page did, and all pass. On ordinary page script it gives Chrome's answer on all 68 expressions of the comparison set and on all 300 further expressions (language, built-ins, async code and the DOM), run against Chromium on the same machine. Three newer sets show what is left. Generators and async ordering match on 206 of 217. Form-control state matches on 197 of 250. Shared typed-array from and of match on 451 of 483. The failures are listed in the harness README.
It does this at a fraction of Chrome's cost: see the [headless engine head-to-head](/benchmarks/headless-engine/) for start time, throughput and memory.
What does this not show?
Know the limits
- This is our run of Lexmount's suite. They have not published a Webfold result, and the published scores for the other engines are for older versions (a later recorded run has Moli at 80.2% and Lightpanda at 61.4%).
- Passing the suite is compatibility on its tasks. It is not a claim about every website.
- Not in the engine yet: canvas drawing, WebGL, web workers, WebAssembly and media playback. BigInt now works. On a probe of 117 web-platform globals and behaviours Chrome answers 98 and Webfold 57. Pages that depend on the missing ones need a real Chromium, which Webfold can connect to.
- No protected-site figure is reported. An earlier 25-site result has no retained public result file, so it was withdrawn; do not plan around bypassing bot protection.
Which other benchmarks are there?
- HTML to Markdown conversion Measured
- Document conversion Measured
- Main-content extraction quality Measured
- Web archive extraction Measured
- Footprint of a scrape or crawl Directional
- PII detection against Presidio Measured
- Localisation extraction Directional
- PDF text extraction with OCR Directional
- PDF extraction against other extractors Directional
- LaTeX conversion with BibTeX Directional
- A whole site to Markdown Directional
- Headless engine over CDP Directional
- Search-results parsing Directional
- Folder search Directional
Check it on your own pages.
Run the command-line tool on your content.