Skip to content
Get packs

Jev Ultrafast

A browser agent that picks one action and element per step, using a single model call per decision.

GitHub Download source
Stars
24.1k
Forks
1.8k
Licence
MIT
Language
Python
Last update
5 days ago
Type
AI model / agent

What it is

Jev Ultrafast is a Python browser agent. You give it a goal in plain language, and it reads the visible controls on a web page, chooses an operation such as clicking or typing, and carries it out in Chrome. A small language model writes text only when the chosen operation is typing. The README reports a 7.1-second Google Flights search, but this is a single demo run, not a general benchmark.

  • Run a goal such as finding one-way flights between two cities and stopping when results are visible
  • Point the same policy at a different site and goal, such as opening a specific Wikipedia article
  • Watch the inspector's numbered elements and probabilities to see why the agent chose each action
  • Run the offline tests or local guard checks to verify changes before use

Great for

  • Developers experimenting with fast, low-cost browser automation driven by natural-language goals
  • Researchers studying structured action spaces for web agents
  • People who want to read a compact codebase to understand how a browser agent loop works
  • Anyone who can run Python and has API keys for the required model services

Not ideal for

  • Users who need a finished no-code tool, since setup requires API keys and a command-line workflow
  • Tasks needing shadow DOM, iframes, canvas, file uploads, or pop-up tabs, which the README says are outside this MVP
  • Production workloads needing proven reliability, since the README says its benchmark covers one task on one browser profile
Step-by-step guide: how to use Jev UltrafastInstall, first run, examples, tips and FAQs
Browser Use vs Jev UltrafastJev Ultrafast vs Stagehand

How to get it

  1. 1

    Clone and build from source

    You'll need: Python · uv (Python package manager) · Google Chrome · TypeSafe API key and a text model API key (OpenRouter in the example configuration)

    Clone the repository

    git clone https://github.com/browser-use/jev-ultrafast.git

    Enter the project folder

    cd jev-ultrafast

    Install dependencies

    uv sync

    Create the environment file

    cp .env.example .env
  2. 2

    Run it

    • Add TYPESAFE_API_KEY and TEXT_MODEL_API_KEY to the .env file
    • Run uv run jev to start the local server
    • Open http://127.0.0.1:8766 in a browser
    • Click Start demo, then Run automatically to watch the agent work
  3. 3

    Get the most out of it

    • Use Choose next in the inspector to pause before each action and inspect the decision
    • Run uv run browser-harness --doctor if Chrome fails to connect
    • Use the library API (Agent class) to run your own goals from Python scripts
    • Switch the text model by changing the endpoint and reasoning settings, as the README describes for Gemini, GLM, and DeepSeek
    • Read performance.md for the measurement details behind the reported timings

Downloads

Source code (.zip)main branch · straight from GitHubOfficial website & docsbrowser-use.com

Strengths

  • The README describes one model round trip per decision, which it says cut browser protocol calls from 1,092 to 101 in its tests
  • Code is organized into a few small files, so the agent loop is readable
  • MIT license and a public repository make it free to use and modify

Trade-offs

  • The README's evidence is limited to three repeats of one task, and the authors say it is not a general reliability benchmark
  • Requires paid API keys (TypeSafe and an LLM provider) and a local Chrome with remote debugging enabled
  • Parts of the web, such as shadow roots, frames, and canvas, are explicitly not supported

How it compares

ProjectStarsTypeLevelUpdatedBest for
jev-ultrafast (this)24.1kAI model / agentintermediate5 days agoDevelopers experimenting with fast, low-cost browser automation driven by natural-language goals
browser-use120kLibraryintermediate4 days agoPython developers who want to prototype an AI agent that operates websites
stagehand25.9kLibraryintermediate6 days agoDevelopers building browser agents that must work on sites they do not control

Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.