Are those speed and cost numbers from a fair comparison?

They come from the project's own harness, and the README points at RESULTS.md for methodology rather than presenting the figures standalone. The setup is more careful than a vibe check: tasks run on a local deterministic site whose success codes are per-session random and therefore unguessable, which removes the need for an LLM judge, plus two live Wikipedia tasks. Still one suite, 12 tasks, one driver model.

What happens when Jev is unsure or stuck?

Low confidence, a stuck page or a detected loop returns status escalate with needs_agent true, carrying the page state, ranked candidates and reasons so the caller can fix the blocker and re-invoke. The verification rule matters here: deterministic diffs of url, text and input values decide whether the last step worked, and Jev's own prev_ok answer is only a tie-breaker when those signals disagree.

Can Jev type arbitrary text into a form field?

Not by generating it. Jev cannot produce strings, so when typing is needed it selects from a joint ref|text choice that binds a value to a field in a single pass, which means the literal string still has to come from you through the texts argument. That is a real constraint for anything user-generated or long-form, and it is also what keeps the action space typed instead of free text.

Is it safe to point at a site I am logged into?

The README says not to, and says why: jev-browser executes an LLM's decisions in a real browser and is not a security boundary. Run it with headless Chromium on non-sensitive sites. What you can rely on is the action space, since Jev can only choose declared options, with an invalid action and element combination getting one corrective re-ask rather than an improvised click.

What MCP tools does it expose?

jev_goal for the autonomous loop, jev_step for one decision and action, jev_observe for compact page state, jev_act for direct actions, jev_ask for batched typed questions, jev_open, and jev_status.

How does a single browser step work?

One snapshot via page.evaluate in 2 to 25 ms builds a candidate list with stable refs and input values, then one Jev call returns the action, element, key, field-text binding, a last-step check, and done-detection.

How does it type text if Jev cannot generate it?

Jev chooses from a ref|text joint choice that binds a value to a field in one pass, and literal strings are supplied through the texts parameter, since Jev itself cannot generate text.

CommunityJev Demos & Experiments

Jev Browser — MahmoudAdelbghany

A browser MCP server where Jev picks the next page action from locally ranked candidates in roughly 300 ms without LLM tokens, deterministic state diffs verify the step, and the driving agent steps in only on escalation.

See highlights

Open github.com

Stars: 0
Language: JavaScript
Published by: Community
Author: MahmoudAdelbghany
Last verified: Oct 2026

MahmoudAdelbghany/jev-browser

User repository on GitHub

View on GitHub

Jev-powered browser MCP for LLM agents — ~300ms decisions, no LLM tokens in the loop. Benchmark vs Playwright MCP included.

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

A Jev‑powered browser MCP server for LLM agents. Jev (TypeSafe AI's decision model) picks the next browser action in 300 ms with no LLM tokens; the driving agent only steps in on escalations. Measured against Playwright MCP with the same driver model: 1.5× faster, 1.6× cheaper, same accuracy on a 12-task suite — and the autonomous Jev loop alone runs at 1.8 s and $0.0005 per task with 97% success. Full numbers: RESULTS.md.

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

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Evaluation Jev for Tool calling

Used inJev for Software & DevTools

Languages: JavaScript

Highlights

One Jev call per step folds verification, done detection and the next decision into a single parallel request against one page.evaluate state read.
Seven tools span the range: jev_goal for the autonomous loop, jev_step, jev_observe, jev_act as a deterministic escape hatch, jev_ask, jev_open and jev_status.
Jev cannot generate strings, so typing goes through a ref|text joint choice that binds the literal value to the field in the same pass.
The benchmark serves a local deterministic site whose tasks end in per-session random codes, so results need no LLM judge, plus two live Wikipedia tasks.

Watch out

The headline numbers are the author's own benchmark against Playwright MCP with the same driver model on a 12-task suite, reported as 1.5x faster and 1.6x cheaper at the same accuracy, with the autonomous Jev loop alone at about 1.8 s and $0.0005 per task and 97 percent success; methodology is in RESULTS.md rather than the README, so read it before quoting the ratios. Success rates on a 12-task suite are not a general claim about web automation. Requirements are Node 20 or newer, a TypeSafe API key with Jev early access, and Chromium installed through Playwright. Two structural limits are worth internalizing: Jev cannot generate text, so any literal string you type must be passed through the texts argument, and the context budget is tight, with choice options capped at 255 and pages budgeted to around 32k tokens. The security note is the one to read twice: the project executes an LLM's decisions in a real browser and is explicitly not a security boundary, so run it headless against non-sensitive sites. 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-browse

Browser agent where Jev makes every navigation decision and agent-browser performs clicks and typing, stopping when the named goal page is reached or blocked.