ArchestackBest of Open-Source AI: List on GitHub ↗3 stars (opens in a new tab)

#2323rd of 33 in Agent frameworks and SDKs

jev-ultrafast

Browser agent that picks an operation and element per step from an indexed element table

Documentation ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
Stars
22.8k
License
MIT
Last commit
Sep 2026
Language
Python

Missed the latest GitHub check; these facts may be older.

Overview

Python browser agent that reads a numbered table of visible page controls and asks TypeSafe's Jev for an operation (CLICK, TYPE_TEXT, SELECT, SCROLL, WAIT, DONE, BLOCKED) and a target in one request. A small LLM generates text only for TYPE_TEXT. The README reports a Google Flights search in 7.1 s and a 25% median speedup over its earlier version on one task.

Who it is for: Developers building fast, low-latency browser automation agents in Python

Strengths

  • One network request per decision; operation and target heads share the same observed state
  • Default loop uses structured DOM state, not screenshots; browser protocol calls dropped 1,092 to 101
  • Model output never becomes selectors, coordinates or JavaScript; executor rechecks freshness and occlusion
  • Small codebase with a local inspector UI and offline tests; MIT licensed

Weaknesses

  • Requires a TypeSafe API key and a separate text-model key (OpenRouter in the example)
  • Shadow roots, frames, canvas, uploads, pop-up tabs and nested scrolling are unsupported
  • Benchmark is three repeats of one task on one browser profile, not a reliability test
  • Shares the existing Chrome profile and needs remote debugging allowed; no release published

What it needs

  • no GPU
  • Needs TypeSafe API (Jev), OpenRouter or other OpenAI-compatible text model API, Chrome, Browser Harness, uv
  • Models: TypeSafe Jev, inception/mercury-2.5 via OpenRouter, Gemini, GLM, DeepSeek
  • port 8766

Also in Agent frameworks and SDKs

See all 33
Also in Agent frameworks and SDKs
RankProjectScore
1ppt-masterAgent skill that turns documents into natively editable PowerPoint decks Live demo ↗ (opens in a new tab)62.1k starsMIT92 out of 100
2langchainPython framework for building LLM agents and apps from interoperable components147.8k starsMIT88 out of 100
3browser-usePython library that lets an LLM agent control a web browser Live demo ↗ (opens in a new tab)118k starsMIT88 out of 100
#4tradingagentsLangGraph multi-agent framework simulating a trading firm to analyze tickers Live demo ↗ (opens in a new tab)111.2k starsApache-2.087 out of 100
#5copilotkitSDK for adding agent chat, generative UI and shared state to apps38k starsMIT84 out of 100