JEV is blowing up. 13,000+ repos on GitHub — these 228 are the ones you actually need to see.

What open-source JEV integrations have in common, and which file to open first.

Search GitHub for "jev" today (28 September 2026) and you get 13,473 repositories. That is a search total, not a count of JEV projects. Some are unrelated names. Many mention JEV in a README and never call it.

From the ones that actually use JEV, we hand-picked 228: repositories where you can open one file and see the exact line where JEV makes a decision.

That list is Awesome JEV: 228 projects in 10 types, every entry pinned to a fixed commit.

How we picked

We did not read all 13,473. We selected 228 relevant projects ourselves. Each candidate then had to clear three bars:

• 50+ GitHub stars
• A JEV decision we could open in the source, not a name in the README
• A fixed commit, so the file you open is the file we read

Two limits, stated up front. Every entry is source-reviewed; none were independently run by us. And stars belong to the whole repository: LangChain's 146.9K stars are not stars for its JEV classifier.

What 228 integrations have in common

1. JEV picks. Your code acts.

Open Jev Ultrafast (18.2K stars), a browser agent. The code builds the list of legal moves from the page: click, type, select, plus two exits, DONE and BLOCKED. JEV picks one operation and one target from that list. Before anything happens, the answer is validated. The choice has to be in the list, and the probabilities have to sum to 1. If the check fails, the code raises "no action executed." A text model is called only when a field actually needs typing.

Open LiteLLM's JEV router (59.4K stars). It sends one choice question named tier, instructed to "pick the cheapest tier whose models can fully answer this request." The router config decides what each tier means.

The same shape keeps coming back:

• QuantDinger (12.0K) checks evidence, exposure and budget before an order reaches the exchange.
• Agentgateway (5.0K) scores jailbreak and secret-leak risk before gateway policy accepts a request.
• OpenViking (38.5K) asks one relevance question per retrieved memory and falls back to vector scores when evaluation fails.
• Skillranker abstains when confidence is below policy.

The reusable unit is not "an agent that uses JEV." It is a closed list your code wrote, one typed answer, and a rule for what happens when that answer is weak.

2. The request format is spreading faster than the model

31 of the 228 are open-model projects. At least 9 of them keep JEV's request format: state plus named Noul, Choice and Score questions, often on the same /v1/systemone path. They swap the model underneath: LocalJev, Jeff, jevmlx, Laya Server and OpenJEV SGLang among them. Several say plainly that they do not reproduce JEV itself.

Laya (17.1K stars) publishes its weights on Hugging Face and answers typed questions in one forward pass.

In practice, code written against the question format has more than one place to run. Several of these projects are built on that assumption.

3. The stars sit in host repos. The ideas sit in the long tail.

The median entry has 251 stars. 151 of the 228 have fewer than 500. Only 20 pass 10K, and most of those are large hosts that added JEV as one option: LangChain, LiteLLM, AI Hedge Fund (63.7K), Composio.

The stranger ideas are small: a Mario agent that picks only legal moves (355 stars), a drone that chooses maneuvers from range sectors (134), a YouTube extension that skips sponsor reads (91), Postgres functions that turn a JEV judgment into a SQL condition (314).

Sort by stars and you find adapters. The patterns are further down.

4. There are already 20 JEV directories

20 of the 228 entries are other JEV lists with 50+ stars each. We are one more, so we did not want to publish another list of links. Every entry here states what JEV is asked to decide and links to the commit where that happens.

Six jobs, one file each

• Filter a live page: Jev Ultrafast → jev_ultrafast/model.py
• Route to a model: LiteLLM JEV Router → jev_classifier.py
• Plug into a framework: LangChain TypeSafe (binary, categorical and ordered-score questions) → classifier.py
• Run it locally: Laya → weights on Hugging Face, plus a checkpoint router
• Gate a request: Agentgateway → guardrail.ts
• Filter before the large model writes: NewsJack. JEV scores hundreds of headlines, and the agent expands only the short list. → demos/news-desk-dealer

The repo also has six scenario guides (filter content, find documents, choose a model, review output, operate an interface, trim context), each listing nearby projects and the part worth copying.

Ask your agent instead of searching

Searching the catalogue needs no API key. Paste this into Codex, Claude Code or OpenCode:

Install the Awesome JEV skill from BeatAPI/awesome-jev.
Then find projects for my task, and point to the exact file worth reading.
Do not call any paid API.

Or install it yourself:

npx skills add BeatAPI/awesome-jev

Then ask: "I want to filter news with JEV. Find 3 reference projects and name the file in each one worth copying."

---

Catalogue: Awesome JEV on GitHub
Browse and filter all 228: Awesome JEV on BeatAPI

Source-reviewed at fixed commits, not independently run. Star counts were captured on 28 September 2026 and belong to whole repositories.