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

Comments URL: https://news.ycombinator.com/item?id=44169052

Points: 2

# Comments: 0

Easy to forget how long the roots were forming.

This is Lucas, one of the creators of Notte.

Notte is an open-source full-stack framework for web agents, designed to be fast and reliable in production.

Our tech revolves around a perception layer that translates the DOM to natural language, allowing LLMs to reason and act upon it reliably. This allows observation of website states and execution of actions using intuitive natural language commands.

The goal with Notte is to ultimately be production-ready for a wide range of workflows.

Notte can be used via a unified API, and includes unified session management, secure vault (payment, creds, keys, cookies, OTPs), built-in stealth, async agent/browser execution, stateful sessions, agent personas (email, phone, identity vault), as well as built-in retry logic and session replay.

We also have captcha pause + HITL integration, as well as an MCP package, plus some cool features for scrapers (JSON schema-driven scraping helpers). All via Python SDK & API.

Apache-2.0 licensed: https://github.com/nottelabs/notte Docs: https://docs.notte.cc/

We’d love to get feedback on:

Is there anything missing before you’d try or trust it on certain workflows? Any edge cases we might have missed? ...or just any general feedback from the HN community.

Let us know what you think!

— Lucas/Andrea

Comments URL: https://news.ycombinator.com/item?id=44040300

Points: 3

# Comments: 0