User repository on GitHub
Fast, bounded browser agents powered by Jev and agent-browser — typed actions, research, classification, and safe orchestration.
- Forks
- 1
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Tools & CLIs
Jev Agent Browser — forvela
Delegated browser execution for parent agents: Jev picks one bounded typed action from the current page snapshot, agent-browser performs it, and ambiguous or stuck flows escalate back to the parent.
- Stars
- 13
- Language
- JavaScript
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- forvela
- Last verified
- Sep 2026
About this project
Give a parent agent a bounded browser task. Jev chooses the next typed action. agent-browser executes it. Ambiguity, repetition, or a stuck state becomes a structured handoff back to the parent. A real 17-second, read-only walkthrough: Tasks → Text Classification → Most downloads → PyTorch → ProsusAI/finbert → back. Watch the full MP4 demo · Open the original WebM recording. Or run the CLI without a global Jev install:
Quoted from the project's README. Read the full README.
Works forJev for Decision gating
Used inJev for Software & DevTools
Highlights
- Jev chooses from a small validated action set, so the browser never runs model-generated code.
- Built on agent-browser with CDP attach, auto-connect, sessions, pinning, and native browser options preserved.
- Browser work is sequential and bounded by --max-steps, and stale refs, repeated actions, and stuck states stop safely.
- Loads ~/.config/jev/config.json for endpoint, apiKeyEnv, pacing, and maxSteps, and keeps secrets out of decision state.
- A 17-second read-only walkthrough is recorded: Tasks, Text Classification, Most downloads, PyTorch, and ProsusAI/finbert.
Quickstart
bash
npm install -g agent-browser
npm install -g jev-agent-browser
export TYPESAFE_API_KEY=YOUR_API_KEY
jev --url https://huggingface.co/models --goal 'Open Tasks and select Text Classification' --max-steps 6Watch out
MIT-licensed; requires Node.js and the external agent-browser peer tool plus a TypeSafe or OpenRouter key, and classification requests send bounded page observations to the decisions API.
Questions about this entry
Short answers drawn from the linked project or page.
No, it is a delegated executor. A parent agent owns intent, permissions and the bounded task; Jev selects one typed operation and target from the live page state; agent-browser performs the native browser action; the loop returns success or escalates a structured state to the parent when the task is ambiguous, repetitive or blocked.
No. Jev chooses from a small validated action set rather than generating code, and the wrapper adds orchestration, not a second browser engine or a generated-JavaScript layer. It also does not bundle an LLM, infer URLs from raw goals, or parse goals with regex. Input values are finite data the parent supplies, and Jev only selects among them.
Two global packages. agent-browser is an external peer tool that owns the browser engine, and jev-agent-browser owns the decision loop and research tools rather than the engine. For the key, either set a TypeSafe key or point apiKeyEnv in ~/.config/jev/config.json at an OpenRouter key by name, keeping the secret out of that file. npx --package jev-agent-browser jev runs the CLI with no global install.
Commit-type actions are meant to stay behind parent or user confirmation: submit, register, apply and similar. Work is sequential and bounded by --max-steps, and stale refs, repeated actions, malformed decisions and stuck states stop safely. Secret inputs are redacted from decision state, traces, events and handoffs.
Jev loads ~/.config/jev/config.json automatically for the decision endpoint, the environment variable holding the key rather than the secret itself, the model, an optional action delay between successful actions, and maxSteps. Configured tools are exposed to Jev as bounded RUN_TOOL choices.
The parent passes the raw goal plus only values it already knows, such as a URL or query text, and Jev resolves the live target from the current snapshot. Missing values return a structured input-required handoff, and secret inputs are redacted from decision state, traces, events and handoffs.
Four surfaces: browser navigation and actions, bounded research that collects raw evidence from configured pages, classification that sends one typed profile-driven batch to Jev, and agent orchestration that streams JSONL events or returns a structured handoff to the parent.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Text-to-screenshot (Malay Vasa)
X: Google's MediaPipe Decision Maker, tested in the browser
X: Playwright CLI + Jev vs Playwright MCP
X: jev-ultrafast splits every browser decision into operation, target and value. JEV passed 18,000 stars in its…
Other listings of this project
Forks and copies of this project in the directory. This entry is the canonical one.
Command-line tool that gives a parent agent a bounded browser task: Jev picks the next typed action and agent-browser executes it, returning a structured handoff when the page is ambiguous or stuck. Includes a research mode and a Node.js runLoop API.
Jev Tools & CLIs#cli#browser#agents
Community
More Jev resources like this
Control a headed Chromium window by voice: one Jev request per partial transcript decides intent, target, and completeness in roughly 300 ms, while code applies thresholds and builds the action.
Jev Tools & CLIs#community#javascript#browser
A Chrome side panel that turns the current page's WebMCP tool schemas into typed Jev questions, predicting the tool call, its arguments, confidence, and latency as you type.
Jev Tools & CLIs#community#javascript#chrome-extension
Community
sdras/jev-webmcp-extensionJavaScript
sdras/jev-webmcp-extensionJavaScriptUnofficial browser automation where the calling LLM states the outcome and Jev decides each element, action, and value from a Playwright page snapshot. Ships as a library, CLI, and MCP server.
Jev Tools & CLIs#community#javascript#browser
Community
Ying-Kai-Liao/jev-browserJavaScript
Ying-Kai-Liao/jev-browserJavaScriptA Claude Code Stop hook that reads the turn transcript locally and spends one four-question Jev call only when files changed with no passing check since, blocking unverified done claims.
Jev Tools & CLIs#community#javascript#claude-code
Community
valentynkit/jev-belayJavaScript
valentynkit/jev-belayJavaScriptBrowser Use's ultrafast agent: Jev picks an operation and an element from an indexed action space in one request, and a small LLM writes text only when the action is TYPE_TEXT.
Jev Tools & CLIs#community#python#browser
Community
browser-use/jev-ultrafastPython
browser-use/jev-ultrafastPythonAn experimental Codex skill and JavaScript runtime that has Jev pick macOS Accessibility elements and actions from text candidates, while local policy gates hold sensitive operations for human confirmation.
Jev Tools & CLIs#community#javascript#codex
Community
Sac-Y/Jev-cuJavaScript
Sac-Y/Jev-cuJavaScriptFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.