v0.1 · MCP server + CLI · MIT

Browser use for coding agents.
3-5x faster than browser-use.

Claude Code, Codex or any MCP client sends a goal. TypeSafe Jev picks the next click or keystroke in about 300 ms. Your agent supplies the text to type, reads the page, and gets control back when jev-ra cannot continue.

Star on GitHub

Works with Claude Code Codex Cursor Any MCP client Python CLI

8.5xWikipedia lookup2.7 s vs 23.1 s · median of 5

7.5xGoogle Flights one-way search, verified8.9 s vs 66.4 s · median of 5

4.0xE-commerce sort3.8 s vs 15.1 s · median of 5

jev-ra median of 5 runs vs browser-use 0.13 flash_mode with gemini-3-flash, same Chrome, same OpenRouter key, 2026-09-18. The headline stays at 3-5x because even our median against browser-use's fastest single run is 5.8x / 5.6x / 4.6x. Method and raw rows in BENCHMARKS.md.

Demo

A recorded run, replayed in the page

The browser draws this from the log of one Google Flights run made on 2026-09-18. Each click, typed value and timing comes from that log. It is not a video.

Source: demo.js (a dependency-free SVG replayer) fed by demo-run-flights.json. Raw screen recordings of the benchmark tasks are in assets/demo/.

Loop

How a step works

Each step is five operations. The only network call is one request to Jev with typed questions. jev-ra builds no prompt and generates no text, and model output is never used as a selector or as script.

Observe

one Runtime.evaluate runs snapshot.js: visible text, every control as [e12] button "Search", identity guards

1 CDP call

Decide

one Jev request with five typed questions: operation, target, which value, did the last step work, is the goal done

~300 ms

Act

guard re-checked, geometry re-resolved, hit-tested, then a trusted Input.dispatchMouseEvent / insertText

trusted input

Verify

url, title, text and field values diffed before vs after; loop detection on three unchanged pages

deterministic

Finish or return

DONE is accepted only with goal_achieved ≥ 0.6 in the same answer; otherwise the run returns with a reason and the top candidates

8 reasons

browser-useflash_mode, gemini-3-flash

5.8 s/step

jev-raJev via OpenRouter

0.7 s/step

Wall time divided by steps on the Wikipedia task (23.1 s / 4 vs 4.3 s / 6, 2026-09-18). A jev-ra step is one decision of about 300 ms plus the page's own load and settle time.

Request

What one request looks like

Jev is a decision model. jev-ra sends the page as state with five typed questions and gets back a chosen option with probabilities and two yes/no answers with confidence. A step costs about $0.00025 and there is no prose to parse.

request · state + questionsPOST /decisions

operation choice CLICK · TYPE_TEXT · SELECT · SCROLL_DOWN · SCROLL_UP · WAIT · DONE · BLOCKED

click_target choice [e3] link "신상품순" · [e4] link "판매순" · [e9] button "장바구니" · …

value_for_field choice origin: "Zurich" · destination: "London" · none

prev_ok noul Did the previous action have its intended effect on this page?

goal_achieved noul Is every requirement of the goal visibly satisfied on this page?

state: { page: {url, title, text ≤ 6000 chars}, elements[≤ 250], recent_actions[≤ 10], goal, values_available }

answers · 312 mstypesafe/jev-1.13

operation → CLICK p=0.91

click_target → [e3] link "신상품순" p=0.87

value_for_field → none

prev_ok → true 0.94

goal_achieved → false 0.12

→ act: click e3 (guard ok, hit-test ok) · verify: text changed, page_changed=true · next step

Question ids and instruction texts are the ones in jev_ra/decide/questions.py; the answers shown are an example. Values your agent passes are chosen by name; jev-ra does not invent strings to type.

Architecture

Your agent talks to jev-ra over MCP (stdio). jev-ra drives Chrome over the DevTools Protocol and calls Jev over HTTPS once per step.

Why there are no screenshotsThe snapshot is text and an element table, so a step sends a few KB and Jev answers in about 300 ms. A screenshot tool exists for pages where text is not enough.

Why the model never touches the DOMEvery target is an observed node id. Before input, jev-ra re-checks the element's identity guard and hit-tests its geometry; a covered or changed control returns stale instead of a wrong click.

What your agent still doesPlanning, supplying text values, and reading the page. jev-ra executes one concrete goal and returns the page with candidates when it cannot continue.

Python 3.12+ uv · uvx jev-ra Node launcher · npx jev-ra MCP · stdio Chrome DevTools Protocol TypeSafe Jev · jev-1.13 OpenRouter

Escalation

When a run stops early

A run that cannot finish returns one of eight reasons, the top candidates and the page text, so the agent can act without observing again.

needs_valuea field needs a string you did not supply. Add it to values, call again.

stuck_loopthree steps, no page change. Read the text, click one candidate, continue.

unverified_doneJev says done, no visible proof. Check with browser_extract.

stalethe page kept changing under it. browser_wait(), retry once.

budgetsteps or time ran out. Narrow or split the goal.

too_many_controlsover 250 controls in view. Scroll or open the section first.

blockedlogin, captcha, empty results. Tell the user; do not retry blindly.

invalid_decisionan answer not on the page, twice. Re-observe and drive one step yourself.

Benchmarks

jev-ra bench --live reproduces every number below: browser-use and jev-ra in the same Chrome with the same key.

Task	browser-use flash_mode	browser-use default	jev-ra	ratio
Wikipedia: open the Gödel incompleteness article	23.1 s · 4 steps	46.5 s · 5 steps	2.7 s · 2 steps	8.5x
Google Flights ZRH→LON one-way, 7 checks verified	66.4 s · 11 steps	63.6 s · 11 steps	8.9 s · 11 steps	7.5x
Olive Young category: sort by newest	15.1 s · 3 steps	17.5 s · 3 steps	3.8 s · 2 steps	4.0x

jev-ra columns are medians of 5 runs (5/5 success each); browser-use columns are the single baseline runs, and its own 5-run medians are slower (24.6 s / 73.8 s / 27.6 s). BENCHMARKS.md carries p90, cost per task and the per-step time breakdown.

Install

An OpenRouter key is enough; no TypeSafe account is needed. jev-ra doctor checks the key, Chrome and one live decision.

export OPENROUTER_API_KEY=sk-or-...
uvx jev-ra install claude
uvx jev-ra doctor
# then in Claude Code: "search the web for X and summarise"

export OPENROUTER_API_KEY=sk-or-...
uvx jev-ra install codex
uvx jev-ra doctor
# Codex asks you to trust the new MCP server once

export OPENROUTER_API_KEY=sk-or-...
npx -y jev-ra install claude   # or: install codex

{
  "mcpServers": {
    "jev-ra": { "command": "uvx", "args": ["jev-ra", "mcp"], "env": { "OPENROUTER_API_KEY": "sk-or-..." } }
  }
}

uvx jev-ra run https://en.wikipedia.org/wiki/Main_Page \
  "Find and open the article about Gödel's incompleteness theorems" --json
uvx jev-ra search "python 3.12 release date" "the exact release date, with source"
uvx jev-ra doctor

If an agent installs this for you, point it at AGENTS.md: install steps, the tool table, what to do on each escalation, and cost.

Limits

These return an escalation with a reason.

Out of scope in v0.1

Cross-origin iframes (same-origin and open shadow roots work)
Canvas-drawn UIs
File uploads
Captchas, bot walls
Multi-tab workflows

By design

Types only values your agent supplied; personal data is never guessed
Model output is never used as a selector or as JavaScript
Done is reported only after a visible check
Jev receives visible page text, not file bodies or screenshots