Jev for browser agents — a daily feed

I first thought this was an X bug, but then I remembered a Chrome extension I vibe coded last week demonstrating the use of jev. https://t.co/pfj5nyEQuf

A 7‑second Google Flights search isn’t magic—it’s a classification problem, not a generation one. Most browser agents waste a full LLM pass to “write” the next click or keystroke. Jev Ultrafast replaces that with a single‑shot pick‑from‑list, turning every step into a cheap network request. Why it matters • The agent calls the browser only 101 times (vs ≈ 1 k) and still loads the page, clicks, typ

Feeding 30k tokens and 4K screenshots to an LLM just to click a button is broken. Humans don't work like that: • System 2 plans the goal • System 1 clicks by instinct in 100ms So I built jev-fast-playwright: an open-source MCP server that fixes this. 🧵👇 https://t.co/nGKX48yV4K

Can Jev Nav? We put Jev, Astra, Fable, Dimcode, and other models + harnesses to the test on 2000+ navigation tasks in 133 real and simulated environments How far are LMs from zero-shotting complex real-time control tasks? Full dataset and code open source ⤵️ https://t.co/73XulmsSDL

3 WORDS JUST MOGGED A PAID SCRAPING API. the page said "was $493.00". an old price, not the current one. every single model returned 493 as the price. then they added one line to the prompt: "use null for any field whose value is not on the page. do not guess." 1 of 16 still fell for it. the benchmark ran yesterday: 16 models 42 twin-page traps 7 page types 573 missing fields without the line: 70.

NotionのAPIで操作できないボタンや自動化作成をCodex＆Jevで実用レベルまでスピードアップすることに成功。 あと1ヶ月ぐらいでNotionの構築はほぼAIに任せられそうだし、それもスキル化できる。恐ろしいな。 僕が集中するのは目標設計・現状棚卸し・業務設計・社内調整・教育。 https://t.co/tt48D7ekDO

Jev Brain — Launch window opens tomorro We didn’t spend the last 48 hours writing a teaser thread. We shipped the loop end-to-end: • Hourly inference credits accrue from holding `AxwSUUHx6hj8bgdtSxVUiKtKkZwmcDbNbEEtTvzfpump` • Unused credits list on the P2P market, escrowed, priced by you, settled in native SOL • Builders buy compute with Phantom. No token swap required. Credits land in their acco

Android操作を、座標の羅列ではなく観察・判断・検証のループにする。 jev-mobileは、TypeSafe JevとMobile MCPを使う耐久型Androidサブエージェントです。 SQLiteにタスクを保存し、observe → normalize → decide → mutate → verifyの流れで端末を操作します。 https://t.co/ZKcEF0qHSu

This weekend I did something cool! For the @typesafeai and @AICollectiveCo hackathon we made Brailly, a way to read the web in braille with the power of Jev! Thanks to @nicopujia @MartinPulitano and @Julian_Irusta04 for being such a great team. https://t.co/WHVgTKluLc https://t.co/bAayA3rc5F

“JevでXのタイムラインを浄化する拡張機能を作った” https://t.co/H5gwKVWpdi

We have made this crazy fast Job applying skill that uses Jev to fill application with speed 🚄 check it out https://t.co/k2e04W3g37

I gave Jev $100 and one rule. Pay your own bills or I pull the plug. Every night at 00:00 UTC it pays for its own server and its own Claude tokens. If the wallet can't cover it, Jev deletes itself. There is no backup. Day 1 starts now. Jev doesn't trade. It works. Open source projects post cash bounties on GitHub issues. $50 to fix a bug. $300 for a small feature. Paid the moment the PR gets merge

Jev pattern on Solari without Jev's API: ARC gives the model a numbered accessibility tree, and one small call per step picks {op, index}. Gemini 3.8 Flash went 24/24 incl. Google Flights.Vs Opus + Solari's MCP on the same browser: 1.9× faster, 39× fewer tokens, 33× cheaper https://t.co/IefeZFMRTK

Playwrightの操作に、Jevの判断と決定論的な検証を組み合わせる。 tontoko/jev-browserは、実際のページ要素を対象にJevが判断し、Playwrightが操作を実行するブラウザ基盤です。 SDK、永続CLI、MCPサーバー、ネイティブ操作とアサーションを一つのコアで扱えます。 https://t.co/JOECVehUwH

Jev-ultrafast使ってみたんだけど ブラウザ操作かなり速いね いまマネーフォワードと連携する経理アプリ作ってて 開発中のAIに同じ画面を何度も開かせてるんだよね 固定資産台帳とか確定申告の画面とか で、普段よく使うサイトなら Chromeの履歴から目的のURLを選んで 直接開いた方が早くね？って思った それでJev Bookmarksってのを作ってみた AIが目的を渡すと 保存したページやChromeの履歴からJevが行き先を選ぶ 開いた先もJevで操作する感じ 役に立ったページはプロジェクトごとに残せるから 次からはそこから始められる 何度も同じ管理画面や設定ページを開かせてる人には こういう使い方も便利なんじゃないかな 履歴にいいのがなければ そこからJevで探せばいいしね あと使ってて思ったのは 少しでも触ってみないと使えるアイデアに気づけないなってこと 自分の作業に入れて

I tried Jev-ultrafast, and it gets around a browser pretty quickly. I'm building an accounting app that works with Money Forward, so my coding agents keep opening the same pages. Fixed assets, tax returns, that kind of thing. That got me thinking: for sites I use all the time, couldn't we just pick the right URL from Chrome history and open it directly? So I made Jev Bookmarks. Give the AI a goal,

”もとの濁りの田沼恋しき” になったりして（笑） / “JevでXのタイムラインを浄化する拡張機能を作った” https://t.co/CzxPe8iZZd

Jevでサイト作ったー！ ブロック避けるゲーム https://t.co/zR36q1tqE7 お買い物 https://t.co/Utc8fvxydh

My terminal for AI agents drives web pages from plain words, now 3.5x faster. Jev picks each click from what the page offers. No chat model mulling it over. Same form, same goal: 18 seconds before, 5 now. Each move is saved as a script. Run two needs no AI. https://t.co/1DzW6esyWs

AI並行開発ターミナルのブラウザ操作が、Jev対応で約3.5倍速くなりました。 やりたいことを書くと、Jevが次に押す場所を選んでくれて、18秒かかったフォーム入力が5秒で終わります。 操作はスクリプトで残るので、2回目はAIなしで動きます。ぜひ試してください！ https://t.co/g4GEyl7Mvu

jev can now steer a browser on its own, just tell it what you want done. jev-browser-skill runs an isolated playwright chromium browser, letting jev choose each in-page action toward a narrow, defined goal. > works with vercel's ai gateway or typesafe's api > flags sensitive steps for review, not guaranteed > forked from cline's official jev-browser plugin 100% free & open-source.

Add your jev sites and workflows on https://t.co/vK4I9tXbZL A collection of sites using jev and videos of jev use cases

JEV + OPUS 5.5 CAN REDESIGN A LIVE WEBSITE AS YOU SCROLL paste any link, hit Start, and watch each section get rebuilt in real time. JEV decides the layout and style in ~0.1s. Opus 5.5 writes the TSX + CSS. then the new section swaps in instantly. 8 sections of a 2015 hosting site were rebuilt in ~20 seconds. JEV decides. Opus designs.

JEV + Opus 5.5 is not fair for web design… 27 redesigns in 40 seconds. Just by scrolling. Each section gets a new design as it enters the viewport. Here’s the loop: 1. A section scrolls into view. 2. Jev returns a typed decision in ~0.1s: { layout, type, palette, drop, p }. 3. Opus 5.5 builds the section. 4. The new design wipes in as the previous version fades out. 5. The update finishes before t

#jev di @typesafeai mentre naviga in autonomia il sito del Comune di Rimini. https://t.co/yVFFJuXjn1 https://t.co/PL90UcWwk1

JEV + Opus 5.5 is not fair for web design… Every section you scroll through is re-designed - 27 times in 40 seconds How it works: 1. a section scrolls into view 2. Jev returns one typed decision in ~0.1s { layout, type, palette, drop, p } 3. Section written by Opus 5.5 4. new design wipes in, old design fades out 5. next part. no race conditions one at a time pass 1: ink stipple ↓ pass 2: riso zin

I've been building software for 18 years. It’s always nerve wrecking to put my baby out there. I’m forcing myself to do it again today. Meet Bowser: the browser that builds itself. Instead of using extensions Google approved, you make your own mods (short for modifications). You can mod anything including Bowser itself. It's full of surprises. Delightful surprises. Under the hood, it uses Jev and

Jevとkintoneの連携テストプログラムを作成。 ユーザーサポート的なテーブルで、ステータスの変え忘れを防止するため、詳細や対応内容の入力を監視して、文面をJevで評価。 処理中的な内容（調査中ですなど）が入ったら、ステータスを「対応中」に変え、「確認しました」などの内容が入ったら、ステータスを「完了」に変更できるようにしました。 kintoneとJevの連携をしたい！ という方、お気軽にご相談くださいー！

GLiNER2.5-Decide now runs entirely in your browser 🎉 ONNX weights + a WebGPU demo in @nicodotdev's open-jev: 4 typed decisions in ~100 ms on a MacBook, and your text never leaves the tab. Try it: https://t.co/52k03zT0fa

Salesforceの画面遷移と入力のイライラ、すべて終わらせます。 ​Salesforceの操作をサブセカンド（1秒未満）に変革する統合ナビゲーションOS 「furu-agent-bar (v1.0.0)」 をオープンソース（Unmanaged Package）として公開しました！ ​LangChainやTypeSafe AI等の外部AIフレームワークには頼らず、応答速度とコスト・安全性を極限まで高めるため 自前で構築した「JEV（JSON Event Vector）アーキテクチャ」 を搭載。 AIが高速で判断（Choice/Score/Noul）し、Salesforceが確定的にDML実行する役割分離を実現しています。 ​💡 現場の「非ITユーザー」が迷わず使える工夫: ​自然言語＆ラベル完全対応: API名ではなく「直近のHot取引先の商談」で即検索 ​全角/半角・複合和数字パース:

Playwright CLI + Jev is 98% cheaper and TWICE as fast compared Playwright MCP Decided to test out Jev by TypeSafe AI and hook it up to Playwright CLI. I compared the results with Playwright MCP to see how both workflows handle a fuzzy instruction such as: "Create new board, list and card" for my Trello app playground. Playwright MCP does what Playwright MCP does - snapshots, decisions thrown to LL

Project Thanos: ask any web page a question, and everything else snaps out of existence.. made w Jev @typesafeai & @claudeai more detailed working in the thread 🧵 https://t.co/SslACIrdI4

I built jev-browser-wingman: an open-source tool that lets AI coding agents (Claude Code, opencode, etc.) use your real, logged-in Chrome. The big LLM still plans. @typesafeai's small Jev model picks what to click or type into at each step. https://t.co/y0v6rdOwjd

Aimed Jev at a tricky problem: finding the next page element on a webpage. Regex gets you part of the way but the long tail of the internet means supporting an endless string of edge cases. Now a quick 100ms call - DOM candidates in, correct target out - and one of the most brittle parts of web scraping is solved.

许多人 对 最近爆火的 新Ai Jev 没有概念🤨 今天让你们感受一下他的超级速度💪🏼 测试模型 上面 Deepseek v4.1 flash ➕Jev 下面 Deepseek v4.1 flash 题目:打开浏览器， 进入b站，找一个 iPhone 17/18的测评视频，播放，然后全屏 Deepseek ➕Jev 只用了35秒 Deepseek 自己用了1分21秒 工具：用的Hermes studio 省了这么多时间这个速度还不恐怖吗，赶紧用起来，兄 弟们🥹🥹🥹

WebMCP gained the least from Jev (2.6×) and was the only surface where every task passed both ways. Browser automation gained the most (4.3×) and had every failure. Try it on your site: https://t.co/EJpw7XWW3C