Organization repository on GitHub
A fast browser-use skill powered by local LLMs via single-token reflexes. Fast, local-first, zero hallucinations.
- Forks
- 25
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Fast Browser Use
An agent skill that reproduces Jev-style single-step logits over indexed browser controls using local Qwen weights for offline browser use.
- Stars
- 221
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- APUS-AI-Lab
- Last verified
- Oct 2026
About this project
An ultra-fast, local-first "System 1" browser automation engine & Agent Skill for Claude Code, Codex, and Cursor. Powered by local Qwen3.5-9B / Qwen3.5-35B-A3B via MLX or PyTorch (CUDA / CPU). Zero cloud inference, second-level reflexes, zero selector hallucinations. English · 简体中文 · Quickstart · Architecture & Deep Dive · Benchmarks · Agent Skill Setup · Python API. Recording preview (1× original playback) · Measurement Telemetry (JSON) · Performance Benchmarks.
Quoted from the project's README. Read the full README.
Highlights
- README frames the skill as open-source reverse engineering of Jev's System 1 discrete decision paradigm.
- Maps visible elements to vocabulary tokens for single-step scoring instead of free-form selectors.
- Targets 100% local execution with Qwen3.5-9B or 35B-A3B rather than the hosted Jev API.
Watch out
Independent weights and calibration; not TypeSafe-hosted Jev. Smaller star count than cloud browser-use demos.
Questions about this entry
Short answers drawn from the linked project or page.
A reimplementation. The README frames itself as open-source reverse engineering of Jev's System 1 discrete decision paradigm and runs entirely on open Qwen3.5-9B or 35B-A3B weights through MLX or PyTorch, with no TypeSafe endpoint involved. Accuracy and calibration are therefore its own, not Jev's.
Structural within its own action space. A lightweight page scanner inspects the rendered layout tree, keeps only currently visible, interactable elements, and formats them as finite tuples such as (CLICK, btn_search), so the model never writes a selector or Playwright code. That does not guarantee the chosen element is the right one for your intent.
Both, through separate paths. Structural navigation including click, select and scroll uses single-token discrete scoring over the candidate logits, while only a chosen TYPE_TEXT action triggers targeted generative inference to compose the field text. KV-cache broadcasting then evaluates multiple fields or candidates in one batched forward pass.
Partially documented here. The README links a separate performance document and a measurement JSON, but the copy captured for this entry is cut off before the benchmark section, so the comparison table's second-level step latency claim against 3,000 to 8,000 ms for cloud agents is not substantiated by the text available.
Python 3.12 or newer on Linux, Windows or macOS, MIT-licensed, with Qwen3.5-9B or Qwen3.5-35B-A3B run through MLX on Apple Silicon or PyTorch on CUDA and CPU. There are no cloud API calls, no telemetry and no subscription cost; the whole inference loop runs on your machine, and it ships as an agent skill for Claude Code and Codex.
No. The README splits the work explicitly: the host agent is System 2 and handles user intent, multi-file analysis, macro planning and outcome verification, while the local engine is System 1 and handles DOM sensing, micro-action selection, bounded waiting and atomic page interactions. On completion it returns an immutable execution trace and the host verifies the resulting page state before proceeding.
Every legal candidate is mapped to a unique single token in the tokenizer vocabulary, then one forward pass extracts next-token logits and evaluates normalized softmax across candidates. The common page context is prefilled once and the base KV cache is broadcast across candidate dimensions, so several fields or options evaluate in a single batched pass instead of a per-token decode loop.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsHacker News: A one-token logprob wrapper that also classifies webcam frames.
allanrbo151 pointsA single function Jev-like wrapper for LLMs, including vision modelsRead the thread on Hacker NewsMore Jev resources like this
29.3kGitHub stars
The Cua computer-use platform documents a jev-use recipe where TypeSafe Jev picks bounded action IDs and Cua Driver executes clicks and fills in the browser.
Jev Demos & Experiments#community#browser#computer-use
Community
trycua/cuaRust
trycua/cuaRust306GitHub stars
Local macOS computer-use automation on Apple silicon using on-device laya typed decisions and CUA-S1-FORMS for filling, via the Accessibility API.
Jev Demos & Experiments#community#browser#local
Community
FluidInference/FluidUseSwift
FluidInference/FluidUseSwiftThe official Laya playground: eight tabs that show choice, score, and noul decisions on ticket triage, phishing, guardrails, RAG filtering, moderation, routing, and multilingual text.
Jev Demos & ExperimentsPlayground#community#open-models#playground
Community
97GitHub stars
A Playwright-based CodexQA skill where Jev chooses indexed controls on a page-built index while the browser executes the action.
Jev Demos & Experiments#community#browser#playwright
Community
openqa-cn/jev-browserTypeScript
openqa-cn/jev-browserTypeScriptA 400M-parameter Jev-compatible decision model with a browser build: 0.796 on typed decisions against Laya's 0.766, 17 ms per question on an L40S, and an ONNX file that runs in a tab.
Jev Models & Reimplementations#community#open-models#local
Community
A ZeroGPU demo of Julia 1: define a state and any set of choice, score, and noul questions and see per-option probabilities rendered as bars and raw JSON.
Jev Demos & ExperimentsPlayground#community#open-models#playground
Community
From the Jev guides
Original write-ups on what this entry is for.
Browser loops are long series of small decisions, and the expensive model does not need to make any of them. The patterns behind the fastest agents, from element tables to OCR plus accessibility trees.
Jev Tools & CLIs#browser#computer-use#agents
Read article
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.