Jev topic · resources
Jev docs, collections and repositories
Jev grew a documentation ecosystem in a week: a vendor blog, two model listings, 8 competing community lists and more than a hundred repositories. 6 of the pages below are TypeSafe's own, 8 are collections somebody else assembled, and 15 are tools — of which 3 run in a browser with no install and no key. This is what each one is.
Try it without installing anything
Three of the resources below are hosted, need no key and open in a browser — each URL was opened on the date shown. These are the fastest way to see the behaviour this whole topic is about, without reading a repository first.
- featherless-ai/simple-jev Featherless AI Turns any open model into a Jev-shaped classifier endpoint — the cheapest way to test the pattern without buying the API.
- jaredpalmer/kev Jared Palmer Small Jev-like decision models you can train and run yourself (0.8B, 4B, 9B), with the fine-tuning workflow packaged as a skill.
- superagents-lab/jev-search Superagents Lab Web search where Jev picks the sources and the query shape, and a writing model does the answering.
Part of the Jev topic, which carries the specs and the price. Eleven repositories read line by line — including several this star ranking does not surface — are in the field notes.
Start with the vendor's own pages
The launch post and the quickstart are the two pages everything else assumes you have read. Claims on the other four are the maker's own, which is what makes them the ones to check first.
| Resource | By | Stars | Last push | What it is |
|---|---|---|---|---|
| Introducing System One models and Jev | TypeSafe | — | — | The launch post: what a System One model is, why the output is a typed decision rather than text, and the two headline ranges (70–500 ms, 40×–200× cheaper) the product is sold on. The source of both headline ranges — and they are vendor figures rather than measurements. |
| Quickstart | TypeSafe | — | — | The request shape: a state, a set of named questions, and the three answer types (Noul, Choice, Score). |
| TypeSafe evaluations | TypeSafe | — | — | The vendor's own evaluation pages, including the self-consistency study we compare our measured spread against. Carries the 0.0102 mean standard deviation of their 15-repeat study — a different rubric, on different machines, from anything measured here. |
| Jev on OpenRouter | OpenRouter | — | — | The listing our snapshot reads: $0.042 per million input tokens, output free, 32K context. Where every price and window on this site comes from — and the page that moves them. |
| Jev on Cloudflare Workers AI | Cloudflare | — | — | The second place Jev is listed, with the question-type table in Cloudflare's own words. The question types in a third party's words rather than the vendor's, which is why it is worth a second read. |
| typesafe-ai/skills | TypeSafe | — | — | The vendor's own agent skills, installed with `npx skills add typesafe-ai/skills`. The one entry here that TypeSafe maintains as a product rather than as documentation. |
What the community has already collected
Eight lists, all of them younger than a week. They overlap heavily — two of them and you have seen the field — and they differ in emphasis, which is the column worth scanning.
| Resource | By | Stars | Last push | What it is |
|---|---|---|---|---|
| yibie/awesome-jev | yibie | 1,070 | 2026-09-21 | Public projects, integrations and discussion threads. The largest of the community collections by stars. |
| Anil-matcha/awesome-jev-by-typesafe | Anil Matcha | 774 | 2026-09-20 | Use cases, patterns, prompts and starter code, each tied to a source. |
| v-modal/awesome-jev-tools | v-modal | 629 | 2026-09-19 | Tools only, no use cases: the narrowest of the collections and the fastest to scan. |
| AbdelStark/awesome-typesafe-jev | Abdel Stark | 431 | 2026-09-21 | A source-backed field guide to TypeSafe's System One line. |
| wuyoscar/jev-skill | wuyoscar | 389 | 2026-09-21 | Use cases, workflows and agent skills, written to be installed rather than only read. |
| logicrw/awesome-jev-projects | logicrw | 336 | 2026-09-22 | An ecosystem radar of open-source projects, with each entry sourced. |
| cobanov/awesome-jev | cobanov | 318 | 2026-09-22 | Projects built with Jev, each one linked to its source. |
| AnotiaWang/awesome-jev | AnotiaWang | 268 | 2026-09-22 | Applications and integrations for Jev and the System One line. |
Tools you can install
Repositories that put Jev inside something else. The same split recurs across them: a larger model writes, Jev decides, and code owns the loop — five independent implementations drew that line before anybody wrote it down.
| Resource | By | Stars | Last push | What it is |
|---|---|---|---|---|
| browser-use/jev-ultrafast | Browser Use | 16,545 | 2026-09-18 | A browser agent that uses a small LLM to write text and Jev to choose what to do next: a flight search in about seven seconds for about $0.004. The clearest case of the split: its loop, its stop gates and what it refuses to guess are the parts worth copying. |
| tamaratran/fast-jev-compaction | tamaratran | 6,063 | 2026-09-18 | A Claude Code plugin that replaces the compaction summary with Jev decisions about what is worth keeping. |
| jarrodwatts/jev-trader | Jarrod Watts | 1,896 | 2026-09-17 | One trade decision per block on Monad, with Jev reading the order book rather than a model writing prose about it. |
| TianyuCodings/NanoJev | TianyuCodings | 1,870 | 2026-09-21 | A small open reimplementation of the idea: parallel decisions over dynamic candidates, as a readable recipe. Nobody calls the API here — it rebuilds the pattern, which is the clearest sign that the pattern is the product. |
| vinnylarouge/jevlike | vinnylarouge | 1,175 | 2026-09-16 | No description published. |
| droidrun/mobile-jev | Droidrun | 329 | 2026-09-17 | No description published. |
| devagrawal09/jev-review | Dev Agrawal | 505 | 2026-09-17 | A staged code-review workflow and local dashboard, with Jev making the keep-or-drop calls between stages. |
| featherless-ai/simple-jev | Featherless AI | 460 | 2026-09-21 | Turns any open model into a Jev-shaped classifier endpoint — the cheapest way to test the pattern without buying the API. The one to read to see how much of the value is the model and how much is the output contract. |
| kerpopule/hermes-jev-skills | kerpopule | 404 | 2026-09-22 | Routing, memory, compaction and skill selection: the agent's small decisions moved out of the model that writes. |
| jaredpalmer/kev | Jared Palmer | 2,857 | 2026-09-22 | Small Jev-like decision models you can train and run yourself (0.8B, 4B, 9B), with the fine-tuning workflow packaged as a skill. The one to read if you want the decision layer off someone else's API. |
| superagents-lab/jev-search | Superagents Lab | 387 | 2026-09-20 | Web search where Jev picks the sources and the query shape, and a writing model does the answering. |
| wy-coliney/jev-browser-use | wy-coliney | 336 | 2026-09-19 | The split written the other way round: Jev clicks, a larger model thinks between clicks. |
| gargpratyush/jev-router | Garg Pratyush | 317 | 2026-09-19 | Routes each request to the cheapest model that can do the job, using Jev to classify the request first. |
| realZachi/pg-jev | realZachi | 289 | 2026-09-18 | Plain-language questions over Postgres tables, with Jev deciding which of the available queries to run. |
| sutro-sh/jev-align | Sutro | 271 | 2026-09-20 | Builds calibrated decision functions from human feedback, using Jev where a threshold has to be learned rather than chosen. |
github.com API, each publisher's own page, openrouter.ai for the listing Verified 2026-09-22 Star counts and dates read the same day; model facts on the topic cover read 2026-09-18No paid placement Not here yet
Tutorials and videos, because none have been watched — an index of videos nobody has watched is a list of titles. Paywalled write-ups and gated demos, which cannot be checked. And the eleven repositories read line by line, which are in the field notes rather than duplicated here.
Six days old, and moving
Five of the eight collections appeared in the last week, and the largest repository in the list is four days old. Star counts and dates here have a shelf life measured in hours. Most of them were posted in the launch thread first, which is where we found them — that thread and the rest are in what Hacker News made of it.