Organization repository on GitHub
Jev-powered semantic browser use
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Jev Turbo
Browser-use loop where a ~200 ms Jev call picks every step, run twice over the same site and model to isolate what a site map changes.
- Stars
- 2
- Language
- Go
- License
- MIT
- Published by
- Community
- Author
- sightmap
- Last verified
- Oct 2026
About this project
Browser use where a 200 ms model picks every step, measured with and without a map of the site. Give it one goal. A sightmap turns the page into a short list of named actions. Jev, TypeSafe's typed-answer model, picks one and says whether the goal is met. No large model is called to act. jev-turbo runs that loop with the map and without it, over the same site and the same model, and reports what the map changed.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Go
Highlights
- Three exact-variant ikea.com goals go from 10 of 15 reached to 15 of 15, and unsure picks drop from 79 to 2, with the map in place.
- On Google Flights, mapping cuts steps from 15 to 10 and picks under 60% probability from 47 to 4 across 10 runs.
- Jev never writes text: everything the loop types comes from --value or a spec file, and --plan calls a large model once per goal.
- Every run file, per-goal breakdown and score row is committed under bench/, next to a Jev-versus-Claude-Sonnet comparison.
Watch out
It is limited to what the page library can see, so canvas, frames and file uploads are out of range, and capturing a Google Flights page costs 250-350 ms a step. Short goals run 2 to 12 steps on cooperative sites, while the two long journeys cap at 30 and 45 steps and one of them fails both with and without the map because the loop cannot repeat a flow. The figures come from a handful of sites with small repeat counts, and an unsure pick is defined as anything under 60% probability. On a real account, --avoid, which drops matching controls from the pickable set, is described as the only guard. 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.
The same loop runs over the same site and the same model twice, once with a sightmap and once with --no-map, so Jev sees the raw HTML and ARIA tree instead. On saucedemo, whose tree already carries good names, the map changed almost nothing, which the README presents as the honest result rather than hiding it.
The README says --avoid is the only guard on a real account, so use it; it drops matching controls from what Jev can pick. Jev itself never writes text, so anything typed into a field comes from a --value you supplied or a spec file you wrote, and the finish check can be a deterministic condition like view=Cart instead of a model answer.
No. The optional --plan flag calls Claude once per goal to write the spec, and that is the only large-model call in the tool. Everything else in the loop is a single Jev request answering two questions: which candidate element to act on, and whether the goal is met.
Anything the page library cannot see: canvas, frames and file uploads. Short goals run 2 to 12 steps on cooperative sites, while the two long journeys cap at 30 and 45 steps and one fails with and without the map because the loop cannot repeat a flow. A Google Flights page also costs 250-350 ms to capture per step.
npm install -g @sightmap/jev-turbo installs the sightmap CLI with it, or use go install github.com/sightmap/jev-turbo/cmd/jev-turbo@latest. Export TYPESAFE_API_KEY, start a sightmap browser session, then run jev-turbo explore --goal "..." --done-when view=Cart --value user=alice.
It is a deterministic finish check that accepts view=NAME, url=SUBSTR, text=SUBSTR, text_absent=SUBSTR, text_count=N:SUBSTR, component=NAME, prop=Comp.name~value, or history_count=N:SUBSTR; repeat it to AND checks. With no --done-when, the loop stops when Jev's own done answer passes 0.85.
No. Jev never writes text, and everything the loop types comes from --value or a spec file. --plan calls Claude once per goal to author the spec, which is the only large-model call in the tool and is optional.
More Jev resources like this
6GitHub stars
Go driver for native macOS apps where Jev picks the next Accessibility action from live app and window state, with code mapping typed answers to tool calls.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
paulsmith/computer-use-jevGo
paulsmith/computer-use-jevGo46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
941GitHub stars
Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
TypeLLM/TypeLLMPython
TypeLLM/TypeLLMPython98GitHub stars
Self-hosted Docker service and web UI that packages Laya structured-decision inference behind a compatible System One endpoint.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
1Panel-dev/laya-serverPython
1Panel-dev/laya-serverPython39GitHub stars
Mac menu-bar voice controller where Jev picks the next action from observed Accessibility controls and Swift executes it in an observe-decide-execute loop.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
ronadin2002/jev-cuaSwift
ronadin2002/jev-cuaSwift55GitHub stars
Toolkit that makes a model behind an OpenAI-compatible endpoint or running in-process answer typed choice, score, and yes/no questions from next-token probabilities in Jev API shape.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
fidecastro/jevifyPython
fidecastro/jevifyPythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.