JevStyle Browser Agent · Preview

A goal. A browser.
One decision at a time.

JevStyle scores each next action and page element. Choose JevStyle-only decisions or a local generative model that reviews every step, then inspect what the browser actually did.

Explore recorded runs Run it yourself

Measured optimization · separate live runs

Skip review when the page has changed

Verified flight search

485.193 s → 323.359 s: 33.35% less elapsed time in these two runs.

Both used the same Zürich to London goal, February 1, 2027 departure date and local 27B model. The new run passed all nine checks with 21 JevStyle scoring calls, 8 observations discarded before review, 13 reviews and 13 executed actions. Every executed action was reviewed.

Most earlier time was spent processing the 27B reviewer’s prompts. The runtime now checks for page changes after JevStyle scoring and can skip an already outdated review. Consecutive skips are bounded; target freshness and occlusion checks still run before input.

Decision-loop component	Earlier run	Optimized run
JevStyle scoring	50.778 s	42.656 s
27B review	424.502 s	270.622 s
Field text	5.221 s	4.565 s
Browser and other loop work	4.692 s	5.516 s

The later run used AC power with Low Power Mode still reported on; power state was not recorded for the earlier run. Browser/network conditions and caches were live. This is not a controlled speedup estimate and does not compare directly with the upstream 7.1-second system, which used different model services without per-step 27B review.

Timing and provenance ↓ · Optimized flight report ↓

Verified results from the optimized Zürich to London flight search

Actual final snapshot, captured after the loop during independent verification. No flight was selected or booked.

Earlier 485-second flight capture

Historical browser snapshots · 485-second baseline

Earlier Zürich to London run

Still frames

485.193 seconds in the decision loop, 23 JevStyle scoring calls, 23 generative reviews and 13 executed actions. Reviewer: qwen3.8-27b-uncensored-q4_k_m. All nine flight checks passed. No flight was selected or booked. These are discrete recorded snapshots, not continuous video.

Flight snapshotFlight capture report ↓

Actual Google Flights results from the independently verified Zürich to London one-way search

The final frame was captured during independent verification after the loop; it has no invented exact frame timestamp.

Earlier 4B hotel capture

Actual browser snapshots

A hotel search, step by step

Still frames

A separate verified run of the synthetic hotel fixture: 20.54 s, 7 JevStyle scoring calls and 7 local reviews. Reviewer: qwen3:4b-instruct-2507-q4_K_M. These are snapshots after individual actions, not continuous video. Its timing is separate from the suite below.

Recorded frameCapture report ↓

Actual browser snapshot of the synthetic hotel fixture; the verified Casa Flora result is open

Inspect the evidence

Follow every choice

Recorded replay

These are saved runs, not a live connection to your browser. Playback pauses between decisions for inspection; reported timings are from the original run.

8 named scenarios passed across 2 separate release-validation batches using identical executable source. This is not one timed suite. Batches: 27B optimized flight check; 27B optimized seven-scenario regression. The flight task stopped before selecting or booking.

27B reviewed eight-scenario release8 / 8 independently verifiedJevStyle scoring + review of every decisionReviewer: qwen3.8-27b-uncensored-q4_k_m

27B optimized flight check1 / 1 independently verifiedEvery executed action reviewed; changed observations may skip reviewReviewer: qwen3.8-27b-uncensored-q4_k_m

27B optimized seven-scenario regression7 / 7 independently verifiedEvery executed action reviewed; changed observations may skip reviewReviewer: qwen3.8-27b-uncensored-q4_k_m

These are successive development runs, with changes to prompts and stopping policies. The counts are observed outcomes, not a controlled comparison of model accuracy.

Loading recorded runs…

Open your run report

Replay track

No step selected—

Executed action

—

Run traces will appear here.

JevStyle operation scores

JevStyle target scores

Run records and measurement scope

Completion is checked separately from the agent's DONE decision. Runs marked “Independent verifier” are stopped when deterministic task checks pass; the model need not choose DONE. A reviewed or hybrid result includes the generative model and must not be interpreted as JevStyle-only accuracy. A successful fixture run shows one controlled task; it does not establish reliability across arbitrary websites. Failed and blocked runs remain visible.

See the report's timing scope for setup and verification boundaries. These measurements are not a direct comparison with the upstream 7.1-second video.

Scenario	Attempt	Result	Elapsed
Find a hotel	27B optimized seven-scenario regression	passed	63.08 s
Open a research article	27B optimized seven-scenario regression	passed	7.08 s
Search and filter a catalog	27B optimized seven-scenario regression	passed	74.27 s
Submit a synthetic support request	27B optimized seven-scenario regression	passed	69.98 s
Navigate and scroll a guide	27B optimized seven-scenario regression	passed	17.06 s
Select an asynchronous suggestion	27B optimized seven-scenario regression	passed	22.69 s
Find a Wikipedia article	27B optimized seven-scenario regression	passed	67.59 s
Find a hotel	27B reviewed eight-scenario release	passed	57.79 s
Open a research article	27B reviewed eight-scenario release	passed	6.49 s
Search and filter a catalog	27B reviewed eight-scenario release	passed	67.53 s
Submit a synthetic support request	27B reviewed eight-scenario release	passed	62.54 s
Navigate and scroll a guide	27B reviewed eight-scenario release	passed	14.98 s
Select an asynchronous suggestion	27B reviewed eight-scenario release	passed	24.96 s
Find a Wikipedia article	27B reviewed eight-scenario release	passed	74.15 s
Search future flights	27B reviewed eight-scenario release	passed	485.19 s
Open a research article	Model-only stopping pilot	failed	31.43 s
Navigate and scroll a guide	Model-only stopping pilot	failed	22.85 s
Open a research article	Verifier stopping pilot	passed	2.16 s
Navigate and scroll a guide	Verifier stopping pilot	passed	1.43 s
Find a hotel	JevStyle-only full suite	failed	4.55 s
Open a research article	JevStyle-only full suite	passed	2.17 s
Search and filter a catalog	JevStyle-only full suite	failed	56.14 s
Submit a synthetic support request	JevStyle-only full suite	failed	1.93 s
Navigate and scroll a guide	JevStyle-only full suite	passed	1.45 s
Select an asynchronous suggestion	JevStyle-only full suite	passed	4.06 s
Find a Wikipedia article	JevStyle-only full suite	passed	12.07 s
Search future flights	JevStyle-only full suite	failed	50.78 s
Find a hotel	JevStyle feedback attempt	failed	74.07 s
Search and filter a catalog	JevStyle feedback attempt	failed	54.10 s
Submit a synthetic support request	JevStyle feedback attempt	blocked	9.18 s
Search future flights	JevStyle feedback attempt	failed	74.73 s
Find a hotel	JevStyle completion gating attempt	failed	46.36 s
Search and filter a catalog	JevStyle completion gating attempt	failed	51.89 s
Find a hotel	Conditional recovery attempt	failed	47.00 s
Open a research article	Conditional recovery attempt	passed	3.04 s
Search and filter a catalog	Conditional recovery attempt	failed	152.75 s
Submit a synthetic support request	Conditional recovery attempt	failed	78.84 s
Navigate and scroll a guide	Conditional recovery attempt	passed	3.08 s
Select an asynchronous suggestion	Conditional recovery attempt	passed	3.50 s
Find a Wikipedia article	Conditional recovery attempt	passed	25.25 s
Find a hotel	Early per-step review attempt	passed	26.12 s
Open a research article	Early per-step review attempt	passed	2.67 s
Search and filter a catalog	Early per-step review attempt	failed	114.29 s
Submit a synthetic support request	Early per-step review attempt	passed	36.73 s
Navigate and scroll a guide	Early per-step review attempt	passed	6.78 s
Select an asynchronous suggestion	Early per-step review attempt	blocked	20.56 s
Find a Wikipedia article	Early per-step review attempt	passed	24.13 s
Search future flights	Early per-step review attempt	failed	110.48 s
Search and filter a catalog	No-op filtering attempt	passed	50.69 s
Select an asynchronous suggestion	No-op filtering attempt	passed	15.13 s
Search future flights	No-op filtering attempt	failed	178.15 s
Search future flights	Flight review attempt 7	failed	56.81 s
Search future flights	Flight constrained-action attempt 8	failed	81.03 s
Find a hotel	4B reviewed seven-scenario release	passed	18.77 s
Open a research article	4B reviewed seven-scenario release	passed	1.71 s
Search and filter a catalog	4B reviewed seven-scenario release	passed	83.37 s
Submit a synthetic support request	4B reviewed seven-scenario release	passed	16.32 s
Navigate and scroll a guide	4B reviewed seven-scenario release	passed	3.61 s
Select an asynchronous suggestion	4B reviewed seven-scenario release	passed	6.66 s
Find a Wikipedia article	4B reviewed seven-scenario release	passed	19.62 s
Search future flights	27B flight attempt 10	failed	303.49 s
Search future flights	27B optimized flight check	passed	323.36 s

Download the published run report

Published development batches and rechecks

Interrupted batches without a completed outcome are retained without assigning them a success score. The interrupted r13 flight attempt and its later corrected-verifier recheck are separate records; the recheck is not counted as a completed run. The short-ID speed experiment was interrupted after an incorrect trajectory; its code was withdrawn and it is not counted as a successful outcome.

Batch	Batch status	Recorded outcomes
Model-only stopping pilot	Earlier report	2
Verifier stopping pilot	Earlier report	2
JevStyle-only full suite	Earlier report	8
JevStyle feedback attempt	Earlier report	4
JevStyle completion gating attempt	Earlier report	2
Conditional recovery attempt	interrupted	7
Early per-step review attempt	completed	8
No-op filtering attempt	completed	3
Flight review attempt 7	completed	1
Flight constrained-action attempt 8	completed	1
4B reviewed seven-scenario release	completed	7
27B flight attempt 10	completed	1
27B interrupted flight attempt 11	interrupted	0
27B interrupted flight attempt 12	interrupted	0
27B interrupted flight attempt 13	interrupted	0
27B reviewed eight-scenario release	completed	8
Short-ID speed diagnostic (withdrawn)	interrupted	0
27B optimized flight check	completed	1
27B optimized seven-scenario regression	completed	7

Interrupted 27B flight attempt 13: corrected-verifier recheck

How it works

Small choices.
Visible actions.

Observe
Read visible controls and assign each supported element an index.
Decide
JevStyle scores the next operation and compatible targets in one request. The reviewed configuration asks the selected local generative model to review each action. If the page changed during scoring, the new runtime can discard that observation before review; it still checks the chosen target immediately before input. Each run identifies the model actually called. Conditional recovery is also available as a separate setting.
Act and check
Execute the selected action. A separate text model supplies field values when needed. Read the updated page before continuing.

Your machine, your browser

Run the agent yourself

Use Python 3.12+, uv, and Chrome. The reviewed releases use jev-style-2b-decision-v4 for scores and qwen3.8-27b-uncensored-q4_k_m for per-step review and required field text. Earlier runs keep their own model identities and settings.

Download the local runner ↓Set up a local JevStyle model

Browser runtime adapted from Browser Use under the MIT license. The download includes the original LICENSE and the attribution NOTICE.

unzip jevstyle-browser.zip
cd jevstyle-browser
uv sync
cp .env.example .env
# Start JevStyle separately: jev-style serve --release 2b
# Start your local generative model server and configure .env.
uv run jev

Open http://127.0.0.1:8766 for the live local inspector. Credentials stay in your local configuration. With a hosted model, page observations needed for decisions are sent to that provider. A local decision model runs on your machine; text and review follow their configured provider.

Reproduce the reviewed configuration

The download defaults to review off, preserving the JevStyle-only baseline. The following are the tested overrides for the existing local 27B server. Change the loopback endpoint to match your own server; the model alias must match its /v1/models response.

JEVSTYLE_BASE_URL=http://127.0.0.1:8765
JEVSTYLE_MODEL=jev-style-2b-decision-v4
TEXT_MODEL_BASE_URL=http://127.0.0.1:18768/v1
TEXT_MODEL_API_KEY=local
TEXT_MODEL=qwen3.8-27b-uncensored-q4_k_m
TEXT_MODEL_REASONING=none
JEVSTYLE_FALLBACK=1
JEVSTYLE_FALLBACK_THRESHOLD=1
JEVSTYLE_FALLBACK_FORMAT=action_enum
JEVSTYLE_HTTP_TIMEOUT_SECONDS=120

With these values in .env, run:

uv run --env-file .env python scripts/run_scenarios.py --max-seconds 1200 --record-frames --output artifacts/reviewed-eight

The scope is six synthetic fixtures, Wikipedia and Google Flights. The flight scenario searches matching options and stops before selecting or booking. Generative review is part of the measured system. The historical full suite passed 8/8; newer release checks identify their separate batches below. These are not JevStyle-only results or a claim of arbitrary website reliability.

The historical 4B seven-scenario release used qwen3:4b-instruct-2507-q4_K_M through Ollama, json_schema review and a 300-second task budget. Its flight attempts failed. Those records remain visible rather than being relabeled as 27B results.

Review at threshold 0.45 is a separate conditional-recovery configuration. All displayed probabilities come from JevStyle, including when a reviewer selects a different action. See the downloaded README for input budgets and uv run python scripts/run_scenarios.py --help for options. You can open its exported runs.json above; local report files are not uploaded.

JevStyle is independently maintained and is not affiliated with TypeSafe or Browser Use.