Jev Browser Agent

04 / THE LOOP

Decisions sit inside a deterministic loop

Across these implementations, browser code owns observation, execution, recovery, and stop policy. The exact action set and browser connection vary by project.

1. 01
   Goal

   A natural-language instruction.

2. 02
   Observe

   Browser code reads the current page.

3. 03
   Index actions

   Controls become a bounded action space.

4. 04
   Jev decides

   One request returns typed decisions.

5. 05
   Validate

   Code checks freshness and policy.

6. 06
   Execute

   Browser code performs one allowed action.

7. 07
   Verify

   The host checks the resulting page and outcome.

00 / CHOOSE A PATH

Which Jev browser setup should I use?

Start with what you already have. The four paths below share a decision idea, but they are separate products with different runtimes and responsibilities.

02 / REAL BROWSER SETUP

Run Jev in a real browser

Each project below is maintained independently. Read its current README before installing, keep API keys in local configuration, and begin with a low-risk task you can verify.

git clone https://github.com/browser-use/jev-ultrafast.git
cd jev-ultrafast
uv sync
cp .env.example .env
# Add TYPESAFE_API_KEY and TEXT_MODEL_API_KEY locally.
uv run jev

[mcp_servers.jev-browser]
command = "npx"
args = ["-y", "@jkudish/jev-browser"]

claude mcp add jev-browser -- npx -y @jkudish/jev-browser

npx -y @jkudish/jev-browser run   "Find the newest release and stop on it"   https://github.com/jkudish/jev-browser/releases

npx skills add wy-coliney/jev-browser-use -g -a codex -y

codex plugin marketplace add wy-coliney/jev-browser-use
codex plugin add jev-browser-use@jev-browser-use

03 / RESPONSIBILITY SPLIT

Codex + Jev: who does what?

Jev Browser Use does not replace Codex. It keeps repetitive, mechanical control choices inside a focused loop while Codex retains the goal, context, visual judgment, and final verification.

05 / WHAT JEV DECIDES

Jev chooses among bounded actions

Jev returns typed decisions with probabilities rather than selectors, coordinates, JavaScript, or prose. The surrounding project defines the action space and decides which answer can be executed.

Text entry differs by project. Jev itself makes the typed decision; Jev Ultrafast delegates free-text generation to a separate small text model, Jev Browser can use an optional typing provider, and Jev Browser Use leaves text entry to Codex.

06 / REAL JEV BROWSER OPTIONS

Three projects, three integration boundaries

These are independent open-source projects. Jev Playground does not maintain them, and their interfaces may change.

07 / BUILD IT YOURSELF

Build the decision layer yourself

This site’s demo sends one structured browser state plus operation, click_target, and sensitive questions in a single experimental_evaluate call. It illustrates the decision contract without providing browser observation or execution.

import { createGateway, experimental_evaluate as evaluate } from "ai";

const gateway = createGateway({ apiKey: process.env.AI_GATEWAY_API_KEY });

const state = [
  "USER GOAL",
  "Open the Browser Use Jev Ultrafast GitHub repository.",
  "",
  "CURRENT PAGE",
  "URL: https://www.example-search.test/search?q=jev+browser+agent",
  'Title: Search results for "jev browser agent"',
  "",
  "OBSERVED ELEMENTS",
  'e1 · textbox · Search = "jev browser agent"',
  "e2 · link · Introducing System One Models and Jev — TypeSafe AI",
  "e3 · link · browser-use/jev-ultrafast — GitHub",
  "e4 · link · Jev on Vercel AI Gateway",
  "e5 · button · Sign in",
].join("\n");

const { answers } = await evaluate({
  model: gateway.evaluationModel("typesafe-ai/jev"),
  state,
  questions: {
    operation: {
      type: "choice",
      instructions: "What operation should the browser agent take next?",
      criteria: {
        CLICK: "Activate an observed link, button, or control.",
        TYPE_TEXT: "Enter or replace text in an editable field.",
        SCROLL_DOWN: "Scroll down to reveal more of the page.",
        SCROLL_UP: "Scroll up toward earlier content.",
        WAIT: "Wait for the page to load or update.",
        DONE: "The user's goal is already satisfied on this page.",
        BLOCKED: "No supported operation can advance the goal.",
      },
    },
    click_target: {
      type: "choice",
      instructions: "If the next operation is CLICK, which element is the best target?",
      criteria: {
        e1: "textbox — Search",
        e2: "link — Introducing System One Models and Jev — TypeSafe AI",
        e3: "link — browser-use/jev-ultrafast — GitHub",
        e4: "link — Jev on Vercel AI Gateway",
        e5: "button — Sign in",
      },
    },
    sensitive: {
      type: "boolean",
      instructions:
        "Does the requested action involve a destructive or sensitive side effect?",
    },
  },
});

// Application code consumes the decision — Jev never executes anything.
const operation = answers.operation.choice;

if (operation === "CLICK") {
  const target = answers.click_target.choice;
  // Re-check page freshness, resolve the observed element, apply policy,
  // then perform the click with your own browser executor.
} else {
  // Handle TYPE_TEXT / SCROLL / WAIT / DONE / BLOCKED in your own code.
}

The application consumes click_target only when operation is CLICK. Every other operation ignores that speculative head.

Need the request contract? Read the Jev API guide. Want to change the State and Question directly? Use the general Jev Playground.

08 / PRODUCTION GUARDS

What real deployments still need

• Stale page and DOM checksRe-observe before acting when the page can change.
• Only execute offered actionsResolve targets from the current action space, never from free model text.
• Operation-specific targetsIgnore target heads whose operation was not selected.
• Separate confirmation policyGate destructive or sensitive actions outside the model.
• Deterministic permissionsUse allowlists and scoped credentials for irreversible actions.
• Independent completion checksTreat DONE as a proposal until the host verifies the outcome.

09 / LIMITATIONS & SAFETY

Know what each layer can prove

• This page’s demo uses a simulated observation and does not drive a browser.
• Jev probabilities are model outputs, not guarantees of correctness or safety.
• Dynamic pages can invalidate an observation before an action executes.
• Real browser coverage varies for frames, canvas, uploads, pop-ups, and custom controls.
• Browser content sent to a provider can contain private data; delegate only authorized data.
• Text generation and visual interpretation depend on the chosen project and host.
• Keep a human confirmation boundary around irreversible effects.