Awesome Jev usecases
The entries worth your time, presented by what each one proves. Headline numbers are as reported by the project authors unless marked otherwise. Star counts were read live from GitHub on 30 September 2026; sourcing for the index itself: awesomejev.com, awesome-jev-usecases, awesome-typesafe-jev, and jev.directory.
Flagships
The projects with the widest reach and the strongest measurement stories.
| Project | What it does | Headline |
|---|---|---|
| jev-ultrafast trending 21.4k stars | Browser agent whose action space is rebuilt every step: Jev picks the operation and target, a small LLM writes text only when typing is needed. | Zurich to London flight in 7.1 s for $0.0039, 178 ms median decision |
| fast-jev-compaction new 7.2k stars | Claude Code plugin that replaces summary-based compaction: every tool call and result is judged keep, truncate, or drop in one fast request. | Context stays verbatim, nothing kept is rewritten |
| jev-chat-jarvis new 7.1k stars | Mobile chat copilot for WeChat, QQ, X, and Feishu: reads the visible conversation, proposes candidate replies, fills the input box. You decide to send. Read-only, no hooks. | The largest consumer-facing build so far |
| SemIf trending 4.6k stars | Open reimplementation: a frozen 4B model read at the option logits, producing typed decisions without text generation. Also known as openjev. | Runs in a browser, no waitlist |
| vercel/eve new 5.4k stars | Vercel's open agent framework ships Jev as the default evaluation model in its experimental evaluate path. | First-party framework adoption |
| jev-trader | Market maker on Monad/Kuru: one buy or sell decision per block from live market state. | 81 ms model latency per block |
Everyday tools
Things a non-developer would install or feel the benefit of.
| Project | What it does | Headline |
|---|---|---|
| jevmail new | Gmail triage into Needs reply, Updates, Promos, Sales, and Spam. Read-only, runs locally. | 1,000 emails in about a minute for 3 cents |
| jev-skip | Sponsor-segment probabilities on the YouTube seek bar, scored from the caption track alone, no crowd database. | 77% of SponsorBlock coverage at $0.0008 per video |
| PlotVeil new | Chrome extension that covers YouTube comments which spoil the plot of the video you are watching or another title you protect. | One Noul per comment, batched 20 at a time |
| JevNoiseGate | Android app that filters notifications and SMS by meaning rather than keywords. Only what it explicitly flags as noise is hidden. | Meaning-based, not a keyword list |
| DWIM new | macOS command palette that reads the frontmost app's menu tree and presses the item matching your plain-language request. | Never auto-runs destructive items |
| Unclutter, typesafe-adblock | Browser extensions that judge each DOM element: clutter, cookie banner, promotion, or ad. Per-element decisions instead of blocklists. | Blocklists are lists; this is judgment |
| Jev Wrapped new | Reads a year of a public Telegram channel and scores every post for paid ads, clickbait, and emotional pressure, then draws a shareable card. | A Spotify Wrapped for media hygiene |
| JevPDF new | In-browser PDF viewer that searches by meaning: one Noul per line, highlights everything that answers your query, ranks by probability. | No OCR service, no upload |
Developer and infrastructure
The tools that changed how people build this month.
| Project | What it does | Headline |
|---|---|---|
| jev-oncall new | Alert triage: one call of four typed questions per alert, pages on severity probability, sends the uncertain band to a human with a 15-minute ack window, links duplicates. | p50 418 ms, $0.04 per 1,000 alerts |
| Dub (production) | The link shortener calls Jev in its malicious-link check before creating a short link, gating URLs with a typed verdict. | Real production adoption, not a demo |
| perch new 315 stars | Parses a repository into a method-level call graph, then asks Jev to flag likely defects and check project rules written in plain English. | From the repo-scanning school |
| is-malicious, jev-git | Supply-chain and pre-commit gates: scan source, config, and CI files for suspicious behavior before execution, screen staged diffs for secrets. | Blocks before the code runs |
| jev-engineering new | Deterministic rules first, one Jev gate second, shipped as a hook, an MCP server, and a team policy. Publishes a 300-call adversarial test with honest results. | Blunt injections moved 0 of 30 dangerous commands |
| json-render | Vercel Labs generative UI framework with a Jev composer that selects which components and actions a rendered interface contains. | Generative UI, constrained |
| neo4jev 152 stars | Navigates a Neo4j graph one relationship at a time, with a classifier over neighboring relationships. | Graph traversal as typed decisions |
| jev-shell-history 115 stars | Zsh history autosuggestions ranked by meaning from the current input. | Semantic shell history |
| jev-logtriage new | Loki log batches judged into suppress, watch, review, notify, or page. Low confidence goes to review, nothing executes. | On-call noise, filtered |
Data and SQL
Judgment as a query primitive. This genre appeared and matured within one week.
| Project | What it does | Headline |
|---|---|---|
| pg-jev 378 stars | Natural-language WHERE clauses for PostgreSQL with batches, session caching, and spend caps. No embeddings, no vector column. | 129 rows in ~1 s for $0.0009 |
| duckdb-jev new | DuckDB extension for Jev predicates, classification, and scoring over SQL rows with budgets and telemetry. | 1k rows in about 10 s |
| jevql, jevsearch | psql-shaped CLI for judgment queries on vanilla Postgres, and a Cmd-K site search block that re-ranks keyword hits with Jev. | No extension, no vector database |
| mysql-ailike | MySQL plugin adding an AILIKE predicate for natural-language row filters and semantic joins. | Judgment inside WHERE and ON |
| jev-reranker | JSON-in, JSON-out CLI that ranks candidates and extracts evidence with independent Noul checks. | Retrieval without embeddings |
Edge, mobile, and IoT
Decisions outside the cloud, where latency and cost matter most.
| Project | What it does | Headline |
|---|---|---|
| Jev-IOT new | Telemetry anomaly triage and autonomic remediation for smart meter fleets. | 10M+ meters, sub-150 ms, under $35/month self-reported |
| jev-drone | Camera-only quadrotor where Jev is advisory at 2.5 Hz while safety loops stay in code at 50 Hz. | The split-rate control pattern |
| mobile-jev | Android agent that decides each tap from a prevalidated action list. | Uber route in ~21 s / 9 actions |
| PocketJev new | On-device iPhone decisions with MLX and a small vision model. | Decisions with no API call |
| HA-Jev 69 stars | Home Assistant integration: ask a question about your house, get a probability, choice, or score as an entity. | A decision layer for the home |
Open models
Reproducing the interface, not the weights. Proof that the decision-layer idea travels, and your path to running it yourself.
| Project | Approach | Headline |
|---|---|---|
| SemIf / openjev | Frozen 4B model, option logits read directly, constrained parallel decoding. | Runs in a browser |
| jevlike | One-pass scorer trained to choose among changing text options, with Doom, chess, and Wikispeedia demos. | Train your own |
| Laya new | Apache-2.0 open alternative with three checkpoints. | Runs on a free Colab T4; ~30 ms vs ~302 ms, at 0.590 vs 0.974 accuracy their numbers |
| Kev new | Local Jev-style models with released weights and evaluation suites. | Compare on your own task |
| open-jev-deberta | Open reproduction on DeBERTa-v3-large. | Self-hostable |
| AnyJev new 949 stars | Nokia's research arm ships a Jev-API compatible model you can run on a phone. The most credible corporate validation the format has had outside TypeSafe. | ~15 ms, on device |
| von new 610+ stars | Non-autoregressive System One model, sub-15 ms, with a drop-in-compatible endpoint. | Sub-15 ms claimed |
| jeff new 267 stars | Self-hosted drop-in replacement for the hosted endpoint, built on GliFormer. | Run it yourself |
| JevK5 new 123 stars | Open-weight alternative returning typed decisions with probabilities in one pass. | Open weights |
The open models trade accuracy for control. Read their evaluation suites before pinning production behavior to them: the published comparisons generally favor the hosted model on decision quality, while the open ones win on latency and independence. The wave accelerated sharply in late September, with a phone manufacturer, a 15ms local model, and two self-hosted replacements all landing in the same week. None of them claim to be Jev.
More
Hundreds more entries exist, and the directories refresh daily. These are the ones that earned a place here.
awesomejev.com
The comprehensive index: roughly 1,100 entries across every category, with star counts read straight from GitHub.
awesome-jev-usecases
The evidence-backed list: every entry scored against a rubric, self-reported and independent results labeled.
awesome-typesafe-jev
A field guide with live demos and an unusually honest section on what independent tests found.
jev.directory
Real systems and the exact primitives they send. Copy one into the Playground and run it.
Last updated 30 September 2026.