We put Jev inside the Lightpanda binary and built judge() on top of it. @typesafeai
No more sending a whole page to an LLM just to find out it's a cookie banner.
> Blocked pages, CAPTCHAs, cookie walls, logins, paywalls, error pages, empty searches and pages still loading can all return an HTTP 200
> judge() asks a set of predefined questions and tells you if the page has real content, and what kind: article, listing, product or docs
> One call, ~0.2s
In the video:
> 30 sites, ~14s total
> Lightpanda backs off when rate-limited, accepts cookie walls, and keeps or skips each page
We benchmarked AI-native headless browsers built from scratch. Including our own.
Lightpanda vs Obscura vs Moli vs Cloudflare’s Kitesurf.
We’ve been building Lightpanda for over 3 years. We’re 8,000+ commits ahead of the newer ones. But that means nothing if it doesn’t perform.
So we measured speed, memory, CPU and web API support.
We’re not objective. So we published the methodology. Link in the comments.
Lightpanda agent runs on Le Chonk from @MistralAI 🧡
The Mona Lisa has seen many things, but not this.
lightpanda agent \
--provider mistral \
--model mistral-large-4 \
--task "go to https://t.co/f2eHzZ7Xpj and extract the title, date and dimensions in cm as JSON"
Lightpanda 1.0.1 is out.
> Its MCP server and agent mode now load each site's stylesheets by default, so menus and modals hidden on screen stay hidden from the model.
> Web Locks, Compression Streams and RSA-OAEP in Web Crypto.
> zstd via libcurl 8.23. setImmediate is gone.
> Each MCP HTTP session now runs on its own thread.
https://t.co/jaCLOMaLMY
Agents using Lightpanda now read a page the way a person sees it.
On https://t.co/8teCksCvn8:
> The page the model reads went from 1,710 lines to 566
> 132 KB to 48 KB
Lightpanda's MCP and agent load the site’s stylesheets by default, so anything hidden on screen (closed menu, a collapsed submenu, or a hidden caption) stays hidden from the agent.
Fewer tokens per step and less noise to reason through.
https://t.co/L4JjZnY75P
Lightpanda 1.0 works with Selenium.
That makes Lightpanda the only browser that natively speaks every web automation protocol: CDP, WebDriver and MCP.
> Change one fixture and your existing tests run on it.
> Up to 4.4x faster than headless Chrome and 7.8x less memory.
> Test it and open an issue if your suite breaks on something.
https://t.co/o9BFM9k8u8
Lightpanda 1.0 is out ⚡
The browser we built from scratch for machines is out of beta and ready for production.
> 10K commits
> 1.7M passing Web Platform Tests subtests
The full story: https://t.co/UdxE2F6L8c
Jev and Lightpanda are getting on well.
Jev decides, Lightpanda clicks. We sent them to browse SF hostels with parking on a startup budget 🍜
> 129 agents fit on one laptop vs 12 with Chrome
> 11x less memory than Chrome on a real booking site
> 6.7s per task with Jev vs 14.1s with gemini-3.8-flash
Code in comments
A basic fetch to the NeurIPS proceedings page gets a 200 and zero papers, because the list is built by JavaScript. 🫥
From Python, Lightpanda reads all 5,858 out of the live DOM.
On a 100 quote scrape:
> 3 seconds and 35 MB on Lightpanda
> 6 seconds and 1.2 GB for Selenium + headless Chrome
More here on using Lightpanda with Python: https://t.co/yuFdkGUDWg
Lightpanda 0.4.1 is out
> CORS support, --experimental-features cors to opt in
> nth-* selectors ~140x faster
> Native ad-blocker regex engine + request engine, more rules supported
> --strip-mode for cleaner content extraction
> WebDriver session management
> Full click pointer/mouse sequence + real wheel scroll
> --locale and --timezone
https://t.co/MElEawssTv
Lightpanda clicks and scrolls like a real browser now.
A click fires the full pointer/mouse sequence, so menus that open on mousedown open, and the wheel scrolls the viewport. 🐭📜
Code in comments
https://t.co/QjjlIHAZRi
https://t.co/AW4fEora1n
@clxymox Merci pour ce partage! Je te confirme qu'on a déjà des utilisateurs qui ont déployé at scale 😉 A dispo si tu as des questions par rapport à ton projet
Lightpanda now ships a native ad blocker, compatible with EasyList.
> It enforces 99.8% of EasyList + EasyPrivacy network rules, failing ad and tracker requests before they leave.
> Enable with --adblock-lists.
📣 Python users, you can just pip install lightpanda.
The browser ships in the wheel.
Same 100 quotes off a JavaScript-rendered page:
> Selenium and Chrome need 1121 MB
> Lightpanda needs 39 MB
https://t.co/CXjlgHFUA3
Lightpanda 0.4.0 is out
> Breaking: iframes and workers no longer load by default (restore with --load-resources iframe and --load-resources worker)
> PDF and PNG output: --dump pdf and --dump png render the page's markdown, so text and structure rather than a pixel match. Same for Page.printToPDF and Page.captureScreenshot over CDP
> New Web APIs inc URLPattern, navigator.geolocation and WorkerNavigator
> --http-header sets headers on every request, and --http-version 1.1 pins the connection
> Introducing an orderfile to reduce the memory peak
> WebDriver BiDi support is WIP. More on this soon. We'd love your feedback on what's not working
Release notes here: https://t.co/GkeneAo10x
Back to school, back to basics. We wrote an article on how to actually use Lightpanda. 🧑🏫
Local vs cloud. CLI, CDP, MCP, our native agent, or PandaScript. Easy to get confused about which fits.
Best practice depends on what you're building. An agent acting on your behalf isn't the same as extraction at scale.
How to think it through before you build ⤵️
https://t.co/QbmurWpy12