Jev Browser Agent

Plain-English browser automation on Apify powered by TypeSafe Jev — structured click/fill/scroll decisions in milliseconds, not a full chat model on every step. You send a start URL and an instruction; the Actor runs a real anti-detect browser, handles cookie and consent interstitials as part of the same goal, and returns structured JSON (summary, URLs, step log, billing breakdown).

---

What you get (under the hood, without you wiring it)

Layer	What it does for you
Cloak Browser	Stealth Chromium profile per sessionId — fewer bot blocks on real sites
jev-ultrafast	Official Jev loop: numbered controls [N], one System One call per step
Browser Harness	Reliable CDP execution (click, type, scroll) against the live page
OpenRouter	jev-latest for decisions; small chat model only when Jev picks TYPE_TEXT
Apify session store	Cookies + Cloak profile tarball restored on the next run with the same id
Apify Proxy (optional)	Sticky session id tied to sessionId when you enable residential/datacenter proxy

You describe the task once. The agent does not stop at a consent screen and ask you to continue — consent, blockers, and your goal are one continuous run.

---

What is Jev Browser Agent?

Jev Browser Agent opens a public page, builds an accessibility-style action table from what is on screen, and repeats:

1. Jev chooses the next operation (click, type, scroll, select, done) and which [N] target fits your instruction.
2. The browser executes it over CDP.
3. The Actor logs each step until Jev marks the goal done or step budget is reached.

Reuse sessionId (demo, flights-research, …) to bring back cookies and browser profile from the named store jev-browser-sessions. Use resetSession when you want a clean computer.

Public pages only — no username/password inputs.

---

Who uses Jev Browser Agent?

• QA engineers smoke-testing signup, checkout, or settings flows in English—with a step log in JSON
• Product managers asking “is this flow still working?” without maintaining Playwright scripts
• Agencies handing clients a session id instead of fragile cookie exports
• Researchers revisiting a public dashboard on a schedule with restored cookies
• Support teams reproducing a UI path with a structured step log
• Indie hackers validating a competitor pricing or onboarding path before shipping
• Ops teams filling the same public form every morning with a saved browser profile
• Travel shoppers running multi-step flight or hotel searches with sticky residential proxy
• Localization checks opening the same site from a geo-stable session
• Builders replacing a chat “computer use” demo with an API they can schedule on Apify
• Analysts capturing structured JSON (summary, steps, URLs) after a manual-style browse

---

What can this agent do?

• Multi-step flows from one instruction (forms, search, filters, date pickers, wizards)
• Consent and GDPR banners dismissed as part of reaching your goal (not a separate mode)
• Persistent sessions — log in once on a public site, continue later with the same sessionId
• Sticky proxy — same IP across runs when Apify Proxy is on
• Evidence — structured steps with tool, target, success, and Jev action labels
• API / schedule — same Actor from Console, REST, Zapier, Make, or n8n

Example instructions

• Add two todos on the demo list and confirm both titles appear.
• Search flights from Sofia to Berlin, departure October 5 2026, economy, stop when priced results show.
• Open the pricing page, switch to annual billing, and stop when the price is visible.

---

Why Jev instead of a chat computer-use loop?

Chat agent (typical)	Jev Browser Agent
Large LLM call every tool step	System One structured choice over visible controls
Slow, costly on long flows	Sub-second routing per step in practice
Fragile free-form coordinates	Indexed refs from a live snapshot
Opaque “I clicked something”	Calibrated probabilities and step log in the dataset

Fill text is still generated when Jev selects a field — via OpenRouter chat — but navigation and clicks stay on Jev, not on a general-purpose chat model.

---

Dataset output

One row per run (see Overview in the dataset). Highlights:

Field	What you get
summary	Plain-language outcome
finalUrl	Last page URL
sessionRestored	Whether cookies/profile were loaded
stepCount / durationMs	Steps taken and wall time
steps	Each action (click, fill, …), target ref, success flag, detail label
billing	PPE chargedEvents and estimated total (charged at exit on every run that reached the agent)
error	Message if the session failed or Jev blocked

Full OUTPUT is also written to the default key-value store for the run.

---

How to run

1. Open Jev Browser Agent → Try for free
2. Start URL — public https page (prefill: TodoMVC demo)
3. Instruction — what to accomplish, in English
4. Session id — keep stable to restore cookies; resetSession for a fresh profile
5. Maximum steps — default 15; raise for long flows (e.g. travel search)
6. Proxy — enable Apify residential proxy for strict sites; sticky session uses sessionId
7. Start the run → read the dataset row and OUTPUT in the key-value store

---

Input

Input	Description
startUrl	Public page to open
instruction	Task in natural language
sessionId	Named computer (cookies + Cloak profile persist)
maxItems	Max Jev steps per run
resetSession	Wipe saved profile for this id before start
proxyConfiguration	Optional Apify Proxy (residential prefill on cloud)

---

Pricing

Three lines on your bill:

Line	What it is
Platform usage	platform-usage-micro — charged once at exit for Apify platform usage on this run (compute, proxy, storage). See billing.platformUsageUsd on the dataset row.
Jev input tokens	jev-input-1k-tokens — per 1,000 input tokens to TypeSafe Jev (System One), rounded up ($0.001 / 1K).
LLM text tokens	llm-text-1k-tokens — per 1,000 tokens from OpenRouter fill-text calls, rounded up ($0.001 / 1K).

All three lines are pay-per-event on the Actor (no separate platform-usage invoice line).

Successful runs include billing on the dataset row: platformUsageUsd, chargedEvents, and token counts. Failed or partial runs are billed the same way at exit (platform + tokens consumed).

---

Limitations

• Public data only — no credential or login-wall scraping
• Heavy anti-bot sites may need residential proxy and a dedicated sessionId
• Very large pages may clip snapshot text for Jev’s context
• Jev only acts on observed controls; custom widgets may need scroll or extra steps

---

Image Credit

Image credit: TypeSafe — Introducing System One models and Jev