LaVague Is Good, Actually, If You Turn Off The Part Where It Watches You

LaVague lets you point an LLM at a browser and give it instructions in plain English — fill this form, click that button, scrape this table — and it actually works, which puts it in a fairly small category of AI tooling right now.

The architecture is what you'd expect: an Action Engine backed by LlamaIndex, a world model that reads the page context, and Selenium doing the physical driving. You write an instruction. The model reads the DOM, figures out what to click, generates the action. Repeat until done or until it halts confused at a CAPTCHA, which is everyone's problem and not specifically theirs.

The thing that caught my eye, though, is that telemetry ships heavy and on by default.

Not buried-in-settings heavy. Not one quiet checkbox heavy. The kind of heavy where you clone it, run it, and it's already phoning home before you've read the docs. Which is a choice. A specific, deliberate choice that someone made, presumably because usage data is valuable to a company that just open-sourced its core product and needs to show investors a graph going up.

This is the standard move now — open source the tool, instrument the tool, monetize the graph of who's using it for what. There's nothing technically wrong with it. It's just worth knowing you're a data point before you automate your way through a hundred internal web forms.

Disable it, then use it. It's genuinely impressive. The world where AI just... operates the browser for you, filling in the tedious parts of being alive — that world is here, and LaVague is one of the more coherent entries in it.

Just turn off the telemetry.