Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

jev-browser-local

Runs jev-browser entirely on a local Jev-style engine (Qwen2.5 1.5B decisions, 0.5B typing) on an RTX 2060 SUPER, with no cloud API.

BenchmarkprototypeCustom license
View on GitHub ↗

Overview

jev-browser-local wires Joey Kudish's jev-browser, a browser CLI and MCP server that normally calls TypeSafe's hosted Jev on every step, to a fully local decision engine. The engine is parallel-decisions, the author's local Jev-style parallel constrained decoding, running Qwen2.5-1.5B-Instruct for decisions and Qwen2.5-0.5B-Instruct for typing search queries on an RTX 2060 SUPER. No TypeSafe, OpenRouter or OpenAI calls are made.

A bundled fork of jev-browser 0.4.0 (MIT) adds a warm-browser CDP mode, and the project adds a VRAM guard, measured benchmarks and run traces with screenshots. The headline result is a Wikipedia task, searching for Ristretto and stopping on the article, completed with goal achieved in 3 steps and 5.9 s at roughly zero cost on 18 September 2026. The README notes absolute paths from the development machine must be adjusted, and that the project is independent and not affiliated with TypeSafe. It has a custom licence.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich page action should jev-browser take next? (decided locally)(inferred)—
Yes / noHas the goal been achieved?(inferred)yes · no

How it works

  1. Run jev-browser with the local bridge
  2. Local engine scores actions with constrained decoding
  3. Small local model types search queries
  4. Warm browser executes; traces saved

Strengths

  • No cloud API and roughly zero per-decision cost
  • Runs on a modest consumer GPU
  • Traces and screenshots for each run

Caveats

  • Tiny local models are less capable than Jev
  • Paths and setup tailored to the author's machine
  • Custom licence

At a glance

Type
Benchmark
Status
prototype
Language
Python, TypeScript
Runtime
local GPU engine with jev-browser fork
Has code
yes
License
Custom license

Use cases and integrations

How curated lists describe it

  • Run jev-browser on a fully local JEV-style decision engine (no cloud API). Warm-browser fork, VRAM guard, measured benchmarks, run traces.

    logicrw/awesome-jev-projects

Frequently asked questions

Does this use TypeSafe's Jev?

No, it replaces Jev with a local engine running small Qwen models.

What hardware was used?

An RTX 2060 SUPER.

How well does it work?

The README shows a simple Wikipedia task completed in 3 steps and 5.9 s.

Related projects

  • laya-browser-agent★ 20

    Local browser agent that picks each step's operation and element with a Jev-shaped model such as a fine-tuned Laya, no cloud needed.

    MCP serverbetaPython
  • jeva

    Open 2B browser-agent decision model fine-tuned from MiniCPM5 that serves the System One protocol for jev-ultrafast.

    ProjectresearchPython
  • OpenJev

    OpenJev 1.7B-8B: open decision models distilled for general decisions and browser-agent steps, served behind /v1/systemone.

    ModelbetaPython
  • cua

    Computer-use platform (Cua Spaces, Driver, Lume) with CUA-S1 small decision models and a preview jev-use example for bounded actions.

    ProjectbetaPython
  • jev-chat-windows-laya

    Fork of the jev-chat Windows app adding a key-free local LAYA judgment backend and a DPI-aware screen capture fix.

    ProjectbetaPython
  • jev-ultrafast-ai-ui-tests

    Selector-free pytest UI tests for saucedemo using a jev-ultrafast-style agent run fully locally with laya-browser and a small Ollama model.

    ProjectprototypePython