Show HN: I made web agents reliable with smaller LLMs via natural language

Hey HN!

I built Notte to see if converting DOM into natural language could improve web agent capabilities and make them work reliably with smaller models.

The result was using deep DOM parsing and a semantic abstraction layer to transform websites into structured, navigable maps described in NL. Instead of feeding raw HTML, there is a perception layer that means LLMs don't just click the DOM elements, but understand the intent behind them.

I benchmarked it against other agent frameworks and was pleasantly surprised - faster task completion and increased reliability (all open source with replayable/reproducible code).

Beyond the core tech, I also built out unified session management, stealth features, credentials vault, CAPTCHA HITL + some more cool features all via a single API. Still working out some edge cases with dynamic content, but it's been solid for most real-world tasks.

Github:https://github.com/nottelabs/notte

Benchmarks:https://github.com/nottelabs/open-operator-evals

Docs:https://docs.notte.cc/side/introduction/what-is-notte

Some questions for the HN community: What's your biggest pain point with current web automation? Is there anything required before you’d try or trust it on certain workflows? Anyone else tried preprocessing web content for LLMs? Curious what approaches have worked for you.

Thanks for checking it out.

— Lucas