jev-demo
Zero-build local playground for Jev with browser-agent, intent routing and context-pruning presets, including a Playwright agent.
Overview
Jev Playground is a zero-dependency local web demo, documented in Chinese and English, for sending a state and typed questions to Jev and seeing choices, scores and calibrated probabilities. Presets cover browser operation from accessibility snapshots (next tool, target ref, done or not), intent recognition (ticket dispatch, moderation, intent routing) and agent context pruning (which tool calls and results still deserve a place in context).
A beta playwright-jev-agent mode takes only a task goal, builds the state and questions automatically, and lets Jev decide each step while Playwright drives a real Chrome or Edge browser, with every request, probability, action and screenshot visible. When pages have too many elements, candidate parameters are narrowed by parallel recall (several batches asked in parallel, top results merged and asked again). Four offline demo pages, including a simple and a complex order back office, test whether Jev can decompose a vague business intent. An in-process Playwright backend is reported much faster than spawning the CLI per command.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which tool and target element should the browser agent use next? | — |
| Yes / no | Is the task complete? | — |
| Yes / no | Should this tool call and result stay in context? | — |
How it works
- Choose a preset or give a goal to the Playwright agent
- Build state from accessibility snapshots or text
- Jev answers typed questions, narrowing large candidate sets with parallel recall
- Execute the chosen browser action and repeat until done
Strengths
- No bundler or framework; runs with Node 18+
- Full visibility of each step's request and probabilities
- Offline demo pages for repeatable tests
Caveats
- Playwright agent is beta
- Requires a TypeSafe key for live use
Getting started
`npm install` then `node server.js` and open http://localhost:3000; paste a TypeSafe API key at the top.At a glance
- Type
- App
- Status
- beta
- Language
- JavaScript
- Runtime
- Node 18+ local web app with Playwright
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
A zero-dependency local web demo for TypeSafe's \*\*Jev (System One)\*\* decision model: send a state plus typed questions, get choices, scores and calibrated probabilities back.
logicrw/awesome-jev-projectsTypeSafe Jev（System One 决策模型）零依赖网页体验台：浏览器操作 / 意图识别 / Agent 上下文裁剪三大预设场景，发送状态与类型化问题，拿到带校准概率的结构化答案.
hellogumbo/awesome-jev
Frequently asked questions
What can Jev decide here?
Next browser actions, intent routes and which context items to keep.
How does it handle pages with many elements?
Parallel recall: batches are scored in parallel, the top candidates merged, and Jev decides once more.
Do I need a build step?
No; only playwright-core is installed for the browser agent.
Related projects
RAM-efficient Rust coding-agent harness, listed as using Jev decisions for memory pruning, browser navigation and voice routing.
ServiceproductionRustNode app that controls Chromium by voice, asking Jev a dozen typed questions per partial transcript to act before you finish.
ProjectprototypeJavaScriptEnhancement pack for MiniMax Code CLI on local DGX Spark hardware, including a Jev-style local decision browser agent and context patches.
MCP serverprototypeChrome extension page guide whose small typed decisions (request type, element to click, match or not) use Jev or local Laya.
ProjectbetaJavaScriptClaude Code kit wiring Jev into a tool-call guard, sub-agent sizing, file search, a browser agent, review, belay and compaction.
ProjectbetaPythonOne-command npx installer for a set of Jev-powered agent tools: context compaction, browser agent, prompt enhancer and skill router.
ProjectbetaJavaScript