JEV Browser

JEV is a System One model. It makes typed decisions (which element, which option, did it work) in ~100 ms for fractions of a cent, and each decision comes with calibrated confidence. It does not write text or plan. You plan and supply the data; JEV executes and asks you when it is unsure.

Choose the tool

Choose the browser

• driver: "extension" is the user's own Chrome, with their logins, through the jev extension. Prefer it for sites with accounts, bot checks or CAPTCHAs (travel, shopping, banking), and whenever the user may need to step in.
• driver: "chromium" is jev's own Chrome profile. A headless profile cannot show a CAPTCHA to anyone, so do not use headless for sites with bot protection.
• The default auto uses the extension when it is connected. jev_tabs shows whether it is, and lists the tabs jev works with (the user's other tabs only with all: true). If the user needs the extension and it is not connected, tell them to load it and run jev pair.
• With driver.background on (off by default), extension tabs open inactive and selecting a tab does not bring Chrome to the front, so the user can keep typing in other apps. Pass an explicit tab when working across tabs. When the person has to act in a tab (a CAPTCHA, a login), say why, then bring it forward with jev_tabs select and show: true; windows a site opens and browser dialogs can still need the user too.

Write a task

{
  "goal": "Find the cheapest flight ticket from Almaty to Antalya departing in October 2026",
  "site": "https://flights.example.com",
  "params": {
    "from":   { "value": "Алматы",  "about": "departure city" },
    "to":     { "value": "Анталья", "about": "destination city" },
    "period": { "value": { "from": "2026-10-01", "to": "2026-10-31" }, "about": "departure date" }
  },
  "result": { "select": "min(price)" }
}

• Write goal, about and hints in English: JEV is most accurate in English. Give param values exactly as the site expects them: "Алматы", not a translation.
• Dates go as YYYY-MM-DD or a {from,to} range. A range lets JEV pick the cheapest or earliest day in a low-fare calendar; code does the date math.
• Use booleans for checkboxes.
• Credentials go into params with "secret": true. JEV never sees those values. Never put secrets in goal or hints.
• result makes the task end on the results. With select: "min(price)" or "max(stars)" JEV first sorts the list on the site. Then you get the results list: one line per item (title, price, short texts, link). Read it and pick the answer yourself: skip accessories, spare parts, sponsored or unrelated items, and report the chosen item with its link. result.pages (default 1) reads more pages or "show more" loads first.
• result.extract: "code" parses items in code by result.schema (field types: string, number, money, datetime, date, time, duration, url, boolean) and applies select. Use it for bulk collection over many pages, where the text list would be too long.
• Useful policy fields: confidence (see below), irreversible: "ask" | "allow", allowed_domains, max_steps, budget_usd, max_items, and fill_required: "any" when any option is fine for required choices you give no param for (body type, colour, trim): JEV then picks the option that fits the goal and hints (e.g. the saved card), else the first. Text fields and consents are still asked.
• For goals that span several parts of a site (the account, a list, a payment), pass plan: milestones [{do, done_when}] in English, from what you know about the site (for example {do: "open the unpaid ads", done_when: "the page lists my unpaid ads"}). JEV aims at the current milestone and checks each done_when on the page; without a plan it finds the way itself (progress checks, menus, rollback).
• For forms (posting, booking, sign-up) give the goal and the details as params; the task finds the way in ("Post an ad", "Sell", the account menu), walks multi-step forms, skips optional steps and works through payment or plan dialogs that are part of the goal. Sign-in params that turn out not to be needed (already signed in) do not hold the task back. For long flows you can also split the work: one task to sign in, another to fill the form.
• The task handles common site behaviour on its own:
  • consent banners;
  • fields the site prefilled;
  • autocomplete suggestions;
  • results that open in a new tab (JEV picks the tab and the task moves there);
  • "show more" pagination.

While a task runs

jev_task returns immediately, and the task never blocks you. Questions reach you through:

1. <channel source="jev-browser" task_id=… question_id=…> events, when Claude Code runs with channels;
2. the block "JEV is waiting for your answer" appended to every jev tool result;
3. jev_wait {task_id}, which waits up to 55 s;
4. in Claude Code, jev watch <task_id> run in the background (Bash with run_in_background). It prints the event as JSON and exits on the next question or when the task ends, and that exit wakes you. After you answer, start it again.

Answer with jev_answer. Look first (jev_observe, jev_screenshot) when the question is not obvious.

A task shown as interrupted (after a daemon restart or an update) keeps its progress: jev_control resume continues it in a tab at its last address. Secrets it had not used yet are not kept: its status says which; the task asks for them (missing_param) if a page needs them. A task paused with user_takeover means the user is working in the tab: leave it until they resume it (side panel) or ask them.

Do not answer continue to a blocker you have not resolved: the task will only ask again.

When the task ends, report the result with its evidence URL. jev_result gives the full item list.

Confidence settings

Every JEV decision has a confidence. Above act the task acts on its own; below escalate it asks you; in between it takes a second look.

The settings stack in layers, from lowest to highest precedence:

1. global: jev_settings confidence.*;
2. per domain: domains.<host>.confidence;
3. per task: policy.confidence;
4. live: a thresholds answer or jev_control update.

• Presets: cautious, balanced (default), autonomous.
• Per decision kind: overrides: {"ground.choice": {"act": 0.7, "escalate": 0.3}}. The kinds are assess, subintent, ground, verify, extract.
• escalate: 0 means "never ask just because confidence is low".
• Trials: reversible steps (fields, chips, dropdowns, date pickers, "more filters", sorting, overlays) act on the best candidate down to trial.floor (default 0.25), verify the effect, roll back and try the next, up to trial.tries (default 2). You are asked only after that, and the question lists what was tried. Set policy.confidence.trial (enabled, floor, tries) to change it; submits never use trials.
• A hint you give in answer to a question about one step is bound to that step's param, so it will not confuse other steps. Put hints that apply everywhere into the task's hints.
• If JEV keeps asking about things it gets right, lower the thresholds. If it acts wrongly, raise them. jev calibrate in a terminal suggests values based on past decisions.

Direct control tips

• Refs (e12) stay stable across observations while the element exists. Regions are r3.
• A [BLOCKING] region (consent wall, modal) must be dismissed first. Clicking a covered element returns occluded; options.force clicks anyway.
• For autocompletes, type with options.mode: "keys", then run jev_observe view:"diff" to see the suggestion popup.
• jev_ask builds its state from the visible page. Reference data as `page.elements.e12` or `page.regions.r3`. Ask one judgment per question. JEV cannot count, compare numbers or dates, or generate text; do that yourself.

Setup and troubleshooting

• jev_doctor checks the key, JEV latency, Chrome and the extension. Keys are set in a terminal: jev settings set providers.openrouter.apiKey - (reads stdin).
• Provider: provider = openrouter (model typesafe/jev-1.13) or typesafe (official API, jev-latest).
• Extension: load dist/extension unpacked in chrome://extensions (Developer mode), then run jev pair and enter the code in the side panel.
• Debug UI: jev ui prints a local link that shows every step, JEV's probabilities, screenshots and replay.