Jev showcase: what people are building with Jev

A trading platform with an optional Jev pre-trade decision gate. The integration evaluates order context and records typed checks in a decision timeline; this listing does not validate trading performance.

Checks submitted prompts and proposed shell commands. Jev scores ambiguous cases; local policy maps results to allow, ask or block. Requires a provider key and host hooks.

A coding-agent hook that asks Jev to assess tool-call risk, untrusted instructions and skill files. Coverage and approval behavior depend on the host integration.

A supervision layer for Claude Code and Codex that records edits and command results. Deterministic ledger checks control blocking; optional Jev judgments provide advice and uncertainty-aware completion checks.

Kubernetes incident triage that keeps deterministic rules in code and sends only the judgement calls to a model.

Auto mode for the Pi coding agent, with Jev deciding what may run unattended.

macOS computer use with Jev as the decision layer.

Static analysis of shell scripts, with Jev judging whether a command is dangerous.

A browser agent with an indexed action space: Jev picks the operation and the target element, and a small LLM handles TYPE_TEXT.

Desktop automation that reads OS accessibility trees rather than pixels, with Jev choosing the next control.

A team workspace for people and agents. Jev-powered decisions route conversations to specialists, select review agents and choose the runtime and model for new sessions.

Jev integrations for model routing, memory selection, search and agent workflows. Its routing dashboard exposes model tiers and decision confidence; the published screenshot uses a demo home.

A Codex browser skill: Jev chooses navigation, clicks, toggles and scrolling while Codex handles text input, visual judgment and final verification. Requires a browser connection and a supported Jev API key.

Supervises coding agents using job requirements, worker activity and verification evidence. Jev or another supported decision provider assesses whether to continue, steer, verify or request human input.

A search application that uses Jev or another configured decision model to choose sources and rank results. It returns links, snippets and relevance scores rather than generated answers.

An Android agent with a live studio and action traces. Jev selects operations and observed controls; Mobilerun executes them. Requires your own device and both provider keys. The author demo stops at payment selection.

Voice-controlled browser automation: speech becomes text, Jev selects intent and observed targets, and Playwright executes validated actions. Requires Chrome or Edge speech support and a TypeSafe key.

A browser agent exposed through MCP, CLI and a library. Jev picks actions from page elements; code enforces budgets and stopping rules and returns a trace and screenshot.

A local gateway that asks Jev to choose tools for coding agents while forwarding other work to the configured language model. Supports several client protocols and requires provider access.

Routes Codex model calls using Jev choices for capability, thinking effort and route duration. Includes local fallbacks and decision logs. Requires a compatible Codex setup and provider credentials.

Agent memory with semantic, temporal, causal and entity graph views. A System One controller guides memory organization and retrieval; a language model composes answers from evidence.

Camera-only drone in MuJoCo: the flight controller keeps it stable, and Jev, at 2.5Hz, only decides climb, brake or thread the gap.

Browser-grounded social research using Jev to choose bounded operations and socai CLI to execute them. Requires a configured provider and your own authenticated browser session.

Neo4j navigation with a Choice for the next relationship and a Noul goal check in one Jev call. Includes notebooks and a Streamlit UI; missing-key stand-in answers are explicitly labelled.

A question-driven crawler with MCP and CLI access. A decision model chooses links and relevant passages, returning source text with URLs and heading paths. JavaScript rendering is an optional dependency.

Turns typed Jev questions about home state into Home Assistant sensors and automation responses. Includes Assist command routing and usage budgets. Requires a compatible Home Assistant installation and provider key.

A Codex skill and Python CLI that recommends installed skills using description ranking and excerpt checks. Can return no suitable skill or uncertainty; it does not install or execute the recommendations.

A live multi-agent news pipeline at news.aatf.ai: GLM-5.3-Flash writes, Jev decides what goes in and where.

Describes itself as the first general-purpose System One agent for Jev: control flow made of decisions, not generated text.

One trade decision per Monad block, on Kuru MON-USDC, at roughly 81ms of model latency.

Judges toxic flow, market pressure and mean reversion, then hands off to an existing strategy. It deliberately does not place the order.

Talk to your Mac: local whisper.cpp transcribes, one Jev call decides the command.

Browser automation where an LLM plans and Jev decides each step.

Prunes stale tool calls and results while preserving retained text verbatim. Includes an npm library and a Claude Code function-hook plugin with a built-in-summary fallback.

Code search for coding agents. Jev judges which files and source passages are relevant to a repository question before the coding model reads them.

A PostgreSQL extension for semantic filtering, ranking and classification. Jev evaluates plain-language conditions on rows and returns probabilities, choices or scores.

A CLI and Claude Code skill that combines site crawling, rule checks and Jev page judgments into PDF, XLSX and Markdown audits. Requires a TypeSafe key; optional external data has separate costs.

Staged code review: Jev flags the risky diffs first, the expensive model or a human sees only those.

An MCP server exposing typed judgment tools for verification, screening, reranking, classification and patch review. Requires Node.js and a configured Jev provider key.

Classifies documents and splits packets into page ranges using local LiteParse extraction and hosted Jev decisions. Includes a local app, CLI and optional cloud OCR; local parsing does not make inference offline.

An experimental CLI that selects uncertain examples for labeling and uses GEPA to improve Jev AI Functions. Proposed changes require review; higher training scores do not auto-accept them.

A self-hosted, versioned skill library with MCP access and scoped clients. Optional Jev recommendations rank skills for a task using your own provider key.

A local PR reviewer that uses Jev to prioritize changes and presents natural-language explanations alongside code. Posts nothing to GitHub. Live analysis requires TypeSafe and OpenAI keys; demo explanations are prepared copy.

Fish-style zsh autosuggestions, ranked by Jev instead of by recency.

A Claude Code context filter that judges large tool outputs with Jev or a compatible adapter. Keeps uncertain blocks, caches hidden text for recall, and requires supported function hooks.

Evaluates agent traces with typed checks for tool use, grounding, quality and security. Supports Jev through several providers and other decision backends; requires configured provider access.

Overlays a live Jev meter on any video, scoring every sentence as it plays.

A CLI for semantic predicates, routing, scoring and JSONL filtering. Jev provides typed answers while the tool handles thresholds, uncertain results, schemas and process exit codes. Requires a TypeSafe key.

A workbench for screening datasets with Jev, training a separate model and inspecting saved decisions and evaluations. Its public demo uses local rules and a small statistical model, not live Jev or LLM training.

A Java client and Spring AI integration for Jev judgments, guardrails, RAG post-processing and tool selection. Includes runnable examples and typed request/response models; needs a TypeSafe key.

Reads a supplied web page, asks Jev which blocks answer a question, and returns relevant passages verbatim with heading paths. Designed for coding agents and available as a CLI and Python API.

Video search through MCP tools and an agent skill. Jev screens contact sheets and checks candidate frames to return timed matches; image support is experimental.

An independent agent skill with Python helpers for bulk semantic checks, ranking and evidence selection. Teaches agents to use Jev and then inspect the selected source material.

Test agent skills against tasks in local or remote sandboxes. Inspect traces and choose a grading backend; TypeSafe/Jev is optional and requires its own key.

A Pi extension that replaces the prompt-wide skill catalog with on-demand Jev ranking. Loads matching skill instructions and exposes a lexical fallback when the provider is unavailable.

Entity recognition from natural-language type definitions. Code proposes text windows; Jev nominates and verifies candidates, then resolves boundaries and types. Outputs text offsets, probabilities and decision traces.

A pre-release library combining Jev-compatible text judgments with tabular model predictions. Uses your labeled rows and numeric columns alongside text-derived features. Bring both model backends; the API may change.

A GitHub Action for catalog submissions and pull-request classification. Jev answers policy questions; code applies the rules and updates a templated PR comment. Requires repository configuration and provider access.

A browser form for assembling Jev requests — pick a template, fill it in, send.

Generative UI where the model picks components and props instead of writing JSON, so the output is well-formed by construction.

MCP connector putting Choice, Score and Noul inside Claude Code, Claude Desktop, Codex and Pi, as a single Go binary.

An agent skill for writing and improving programs that call Jev.

Filters JSONL and Parquet training data on quality, relevance and risk at 1,500+ rows a second.

Codebase navigation: at each directory, decide which files are worth descending into.

Probabilistic control flow for Ruby — chance, pick and rate as language-level constructs.

An intent-driven UI runtime: buttons declare what they mean, and the runtime routes it.

An input that turns into a task, event or other UI card. Jev classifies intent and signals; deterministic code extracts dates and values. An offline keyword classifier is also available.

A game agent that reads emulator memory and gives Jev legal options for navigation, menus and battles. The harness handles movement mechanics. Requires a legally obtained game copy and provider access; no ROM is included.

An experimental Chrome extension that asks Jev whether candidate page elements are ads, then highlights or removes them. A learning demo with false positives, not a tracking or malware blocker.

A Chromium T-Rex game with manual play and an optional Jev controller. Jev chooses jump, duck or keep running; browser code handles frame timing, collision geometry and execution. AI mode requires a TypeSafe key.

Plays Super Mario Bros. from structured emulator RAM rather than from frames.

Describe a startup idea; several Jev Score questions, combined in code, return KILL, FIX or SHIP.

Browser 1v1 FPS where every tick decides movement, aim, fire and jump.

A chatbot built on a model that cannot write, by choosing replies one word at a time.

StarCraft II played by picking from the legal action space each step.

Generating level terrain during play, one Choice at a time — possible only because each decision is sub-second.

A word game where everyday phrases become absurd synonyms for you to decode, assembled from picks because Jev cannot write.