See what's there.
Visible controls become an indexed action space. The model works with the page in front of it.
[04]searchboxDestination
SMALL DECISIONS. REAL PROGRESS.
A browser
that chooses.
A little intention. A clear next move.
Give Jev a goal, and
watch it find its way through the web.
MADE OF Typed decisions. Observable
actions.
ONE GOAL. LET'S GO.
Find a Design stay in Lisbon with free cancellation. Open Casa Flora.
forma.example / stays
DEMO
Reading the page10 visible elements
Built with intention.
Bun nativeTypeScriptYour modelsOpen source · MIT
JEV / 001
01 — THE IDEA
The whole web.
One next move.
A page is a set of possibilities. Jev gives each one a name, chooses an action, and checks what actually happened.
Make it specific.
One typed decision pairs an operation with a compatible target. Every choice has a place to go.
TYPE_TEXT[04]
A real next step.
The browser clicks, types, selects and scrolls. The inspector makes the journey visible as it unfolds.
LisbonFIELD UPDATED
Check the result.
Observed evidence must satisfy the checklist. A final review checks the original goal before completion.
Goal reviewed
02 — A WINDOW INTO THE WORK
Every move.
In the open.
The useful part of an agent isn't just where it ends up. It's being able to follow how it got there.
Explore the inspector
DECISION TRAILOBSERVED EVIDENCE
01
Destination enteredTYPE_TEXT [04]
Lisbon
02
Category selectedSELECT [06]
Design
03
Cancellation confirmedCLICK [07]
Checked
04
Casa Flora openedCLICK [08]
Verified
From a goal to a checked result. DONE
Illustrative trace from the walkthrough above.
Stay in the loop.
Run automatically or go step by step. Pause, inspect a decision, and follow the pages visited.
Local. Cloud. Your call.
Connect a System One decision model and an OpenAI-compatible chat helper. Switch connections as you work.
Small enough to know.
A Bun-native TypeScript library with a browser inspector and CLI. Open it up. Make it yours.
03 — BUILT FOR BUILDERS
A few lines.
A new direction.
Bring a goal and your model connections.
Jev brings the loop.
Bun 1.4.3+ · Chromium · Two model connections
# Add Jev to your project
bun add @usex/jev-browser-agent
# Give it a browser
bunx --bun --no-install playwright install chromiumYour next browser agent starts here.
# .env — a System One decision model
TYPESAFE_BASE_URL=http://127.0.0.1:1234/v1
TYPESAFE_MODEL=your-decision-model
TYPESAFE_API_KEY=
# An OpenAI-compatible chat helper
TEXT_MODEL_BASE_URL=http://127.0.0.1:1234/v1
TEXT_MODEL=your-chat-model
TEXT_MODEL_API_KEY=Use local endpoints or your cloud provider's URL and key. The chat helper must support JSON schemas.
import { Agent } from "@usex/jev-browser-agent";
await using agent = await Agent.create(
"https://en.wikipedia.org/wiki/Main_Page",
"Find and open the article about Lisbon.",
);
for await (const state of agent.run()) {
console.log(state.status, state.page.url);
}
Save as agent.ts, then run: bun run agent.ts
OPEN SOURCE. OPEN POSSIBILITIES.