XPath

And as always, what I love about Sanjay's stuff is that this is completely free, and his pitch is that most agents read the HTML and guess a selector that looks right.

This tool, though, generates Playwright locators, XPath, and CSS selectors, then confirms each one matches exactly one element with the Playwright locators verified using real Playwright.

It skips generated IDs and hash class names that break on the next build, anchors locators in lists and tables to nearby unique text, It supports Shadow DOM, nested iframes, and CVG.

So you can ask your agent in plain English to fix a broken locator, validate one, list elements on a form, or build a playwright page object.

And then the extraction layer.

CSS, XPath, RegEx if you want it deterministic, or LLM extraction strategy if you want a model to pull structured fields out.

Deep crawl with BFS, DFS, best first, crash recovery.

An adaptive crawler.

When you use this method, you're essentially asking N8N to visit a webpage and download its HTML content, just like your browser does.

But instead of rendering it visually, N8N grabs the raw code and lets you extract specific pieces of information using CSS selectors or XPath expressions.

Here's where it gets interesting.

Most websites serve their basic content in that initial HTML response.

I used to prefer it this way, but I've learned that this is less scalable and less movable.

So the good thing about JSON-LD, because it's scripting and you can just add it into the body of the DOM, it's not visible on the page, even if you can reference very specific types on the page using speakable markup or there's other special selectors for it using XPath and CSS selector properties in the web page element and web page But I would just keep it simple and just say what would apply site-wide with organization websites, almost universal.

And then I would just focus on very specific page types that are the moving mechanisms and think about how it would, like if you were to go from one page to a thousand of these pages and how that would change all the permutations and then find the common elements of it.

So name is always 100%, URL is always 100%.

Tell it to add logic nodes, reconnect the workflow, and rewire UI elements dynamically.

The benchmarks show it can parse massive nested div soup, trace the XPath, and rewire logic without ever needing a command line interface.

And looking at how Astra handles that DOM manipulation really reveals the core strategic tradeoffs between the major labs right now.

How so?

Instead of writing scripts with Momentic, you could define your scripts as plain English YAML files in your code base.

So say, instead of like an XPath, you could just describe the element, like how you go to an end user.

So that was the product that we've been selling for the past two years.

and like you know today's launch we launched mo our uh which is like kind of what ties into like what we were discussing here today is like what can we envision a world without test scripts like no selenium no playwright not even momentic test script because you know i think the fundamental understanding that we came to over the last two years is that people just want no bugs, they don't want yet another testing tool or framework to build or maintain.