Can the model write arbitrary selectors or run arbitrary JavaScript?

No. It chooses from a fixed set of observed operations: click, type, select, wait, done, or blocked, and Browser Harness executes the selected action over the Chrome DevTools Protocol. The text helper is only allowed to supply a value when the chosen action is TYPE_TEXT, so text generation is never in the same position as control.

Why do I need a second model when Jev is the decision maker?

Because typing a value is a different job from deciding to type it. The text model is consulted only when the chosen action needs text for a field, which keeps Jev as the sole decider. You configure it as an OpenAI-compatible endpoint, so you can point TEXT_MODEL_BASE_URL at a local server rather than a hosted one.

What stops a malicious page from hijacking the agent?

Page text is treated as untrusted data, never as instructions, and the model only picks from candidates already observed on the page rather than writing its own selectors. On top of that the interface binds to 127.0.0.1 only, every API call requires a per-session token, and you can switch to manual step-by-step control to inspect every decision before it executes.

Is there any benchmark showing this works on real sites?

The README publishes none, which is worth knowing before you point it at anything important. What it does document is the pipeline and the checks: page observation converted into bounded candidates, Jev ranking the next operation and target, the harness executing, and the UI recording probabilities, confidence, latency and trace history that you can export as JSON.

How do I install and run Jev Browser?

It requires Python 3.12+ and uv. Run uv sync, copy .env.example to .env and fill TYPESAFE_API_KEY plus the text model settings, then uv run jev and open the printed local URL.

Which operations can Jev choose from?

Only observed operations: click, type, select, wait, done or blocked. The model never emits selectors or executable browser code.

What does the local UI let me inspect?

Screenshots, ranked targets, probabilities, confidence and the full decision trail, plus numbered target overlays, a JSON trace export, optional screenshot recording, and manual or automatic execution.

CommunityJev Demos & Experiments

Jev Browser — KesavanKing

A local Python browser automation UI that turns a URL and a plain-language goal into bounded actions, with Jev picking the next operation from observed candidates and a separate text model used only to fill fields.

See highlights

Open github.com

Stars: 0
Language: Python
Published by: Community
Author: KesavanKing
Last verified: Oct 2026

KesavanKing/jev-browser

User repository on GitHub

View on GitHub

Local browser automation UI that uses TypeSafe Jev to choose bounded page actions and a text model only for field values.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

This is one of several Jev listings for the same project. The canonical entry is Jev Browser — jkudish.

About this project

Jev Browser is a local browser automation UI that turns a URL and a plain-language goal into safe, bounded browser actions. It observes the live page, asks TypeSafe Jev to choose the next operation and target, then executes through Browser Harness. A separate OpenAI-compatible text model is used only when a field needs text. The model never emits selectors or executable browser code.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Routing

Used inJev for Software & DevTools

Languages: Python

Highlights

Three jobs stay separate: Jev picks from observed operations, Browser Harness drives CDP, and the text helper fills fields only.
The model never emits selectors or executable browser code, so a bad decision cannot turn into arbitrary script execution in the page.
The UI shows screenshots, ranked targets, probabilities, confidence and the full decision trail, and you can step manually or let it run automatically.
Safety is local by construction: the UI binds to 127.0.0.1 only and API calls need a per-session token.

Watch out

You need two providers configured even though they do different jobs: a TypeSafe key plus a separate OpenAI-compatible endpoint for TEXT_MODEL, which can point at any chat-completions compatible server and is only consulted when the chosen action is a text entry. The architecture deliberately splits deciding, generating text and controlling the browser, so this is the judgment and UI layer rather than the browser driver itself, which it calls Browser Harness. Being bounded to observed candidates is the point and also the limit: it can only act on what was already detected on the page, so unusual or canvas-heavy interfaces are the likely failure mode. The README publishes no accuracy figures or benchmark, and the checks it lists are lint, pytest, a node --check syntax pass and a build. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-ra

Browser-use layer for CLI coding agents: Jev picks the operation and target element in one round trip per step, exposed over MCP with a CLI, a Python API and an HTTP server.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

brnyxx/jev-raPython

jev-tetris

A Streamlit Tetris where the engine enumerates every legal placement, Jev picks one typed placement ID, and Python verifies rotation, column and exact occupied cells before the piece is allowed to lock.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

MachineLearning-Nerd/jev-tetrisPython

open-jev-typed-decision-engine

Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

intikhab49/open-jev-typed-decision-enginePython

TypeLLM

941

Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

TypeLLM/TypeLLMPython

jev-robot-control

Simulation study that drives one xArm7 apple-to-plate task with Jev, a large reasoning model, and a small model through identical incremental controls.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

openroboto-ai/jev-robot-controlPython

jev-cua

Mac menu-bar voice controller where Jev picks the next action from observed Accessibility controls and Swift executes it in an observe-decide-execute loop.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

ronadin2002/jev-cuaSwift

From the Jev guides

Original write-ups on what this entry is for.

Per-Tick Decisions in Games

Sep 28, 2026

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

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai