Jev GitHub Projects Worth Knowing About

Is Jev itself open source?

No, and getting that wrong is the commonest error in Jev GitHub writeups. The model is closed: no weights, no training data, no architecture beyond prose. Nothing on this page is Jev.

Two open things get conflated constantly. TypeSafe maintains permissively licensed client libraries, which is not the model being open — see the official SDK page. Separately, the community writes agents and plugins calling the hosted service, plus independent models copying its input and output shape. Anything sold as “open-source Jev” is that second thing.

How we picked these

Jev GitHub results rank on stars, and launch-week stars measure novelty. So: more than 400, code doing something a README could not, and a link to Jev that survives reading the repository. Then commits, issues and last push, because three commits with 87 open issues is a different proposition from daily work.

Checked 2026-09-21. All six are permissively licensed.

jev-ultrafast

The most-starred Jev GitHub project by a wide margin, from the Browser Use team. It drops the usual loop where a language model writes an action out as JSON. Each observation becomes a numbered element table; one request picks operation and target together from an offered set of CLICK, TYPE_TEXT, SELECT, scroll, wait, done and blocked. A small language model is invoked only to write text.

Their own measurement: median task time 9.450 down to 7.092 seconds, protocol calls 1,092 down to 101, over six alternating runs both passing three of three — labelled three repeats of one task on one profile rather than a reliability benchmark, which is more restraint than the stars predict. Weigh those 87 open issues against the three commits.

fast-jev-compaction

Second on the Jev GitHub star list, and a Claude Code plugin that takes over context compaction. Instead of a model writing a summary, every tool call and result is scored in one request against a keep threshold, 0.5 by default: above it both stay verbatim, below it the result is truncated, below that the call is dropped. Nothing surviving gets rewritten — a summary is where the literal string you needed goes missing.

It is also the only project here documenting what happens when the decision layer lets you down: it logs a fallback to the built-in summary.

SemIf

Formerly OpenJev, renamed for good reason. It runs typed option scoring on open models on one consumer GPU, with a browser demo. Its README is explicit on the two points most Jev GitHub coverage blurs: the project is independent, not affiliated with or endorsed by TypeSafe, and what it reproduces is the interface pattern — option probabilities read straight out of a model, no answer sentence, no JSON repair — not the Jev model or its undisclosed training. Anywhere you have seen it called an open-source Jev, that was the writeup being careless.

Its most useful output is a comparison, not a product: one frozen 4B model, 21 questions, typed readout against a compact generative baseline. Both arrays came back valid and identical; the generative path took 5.21 times as long. The fixtures and raw timings are committed.

NanoJev and jevlike

Two independent attempts at training the same shape: text plus N options in, one probability per option out, in a single pass.

jevlike is the blunter of the two, and its README is the most candid document in the Jev GitHub ecosystem: 26% on target-disjoint Wikispeedia next-click data against an 8% shuffled control, and a chess checkpoint going 0 wins, 2 draws, 48 losses against Stockfish at its lowest level. Untouched since the day it was created. NanoJev is the one still moving and the fuller reproduction: parallel decisions, dynamic candidates, a training pipeline, 896 episodes holding 17,498 recorded decisions. Neither substitutes for the hosted model, and neither claims to.

jev-review

A staged code-review workflow with a local dashboard, in TypeScript, and the best demonstration in the Jev GitHub set of using the model the way it rewards being used: orchestration stays in ordinary code, the model is only asked bounded questions, and thresholds live in code rather than in a prompt. Risk matrix, then file profiles, evidence selection, mechanism classification, severity scoring, reviewer routing.

It reviews a git diff or scans a whole codebase, wants Node 24 or newer and your own key, and binds the dashboard to localhost. At 161 KB it is the smallest repository here — short enough to read in an afternoon.

awesome-jev and the community lists

This is where the Jev GitHub trail stops being worth following. A search on the name awesome-jev alone returns 60 repositories: several generated, several the same links reordered, one advertising a verified catalogue of 503 projects while holding 38 stars.

Two are worth naming, for opposite reasons. yibie/awesome-jev, 673 stars, is pushed to daily. Anil-matcha/awesome-jev-by-typesafe, 719 stars, is the most-starred — and was created on 2023-05-17, three years before any project on this page existed. That is a renamed repository, and its stars are not evidence about its Jev content. Check the creation date on any Jev GitHub list before trusting its curation.

What none of these solve yet

Read the six together and the same gaps appear in every one. What the Jev GitHub ecosystem has not built, you will be building yourself.

Nobody checks the answers. The three projects calling the hosted service ship no way to learn whether its decisions were right on your data. The three shipping evaluation harnesses score their own models, not Jev. The only independent calibration audit of the hosted model has zero stars. Plenty of code that trusts a confidence number, then, and almost none that tests one — while TypeSafe’s own benchmark puts the model at 61.8% on its invoice workflow, which how it works inside goes through properly.

Nobody handles being wrong. Every call returns a probability and exactly one project does anything with it: fast-jev-compaction, with a documented threshold and a documented fallback. The rest take the top choice as the answer.

Nothing is versioned. Zero published releases across all six. Three have five commits or fewer. Pinning one as a dependency means pinning a commit hash and hoping.

Nobody has touched the ceilings. Long inputs and per-call cost have one experimental repository each, neither above 60 stars. If either constrains you, that is your work, not a package.

None of which argues against the model. It argues for treating every Jev GitHub project as a worked example to read rather than a dependency to adopt.

Frequently asked questions

Is Jev open source?

No. The model is closed and unpublished. Open are TypeSafe’s permissively licensed clients and the tooling written around the hosted service.

Can I run Jev locally?

Not Jev itself. The closest options on Jev GitHub are SemIf, NanoJev and jevlike, and all three reproduce the interface rather than the model — different weights, different training, different accuracy, as SemIf says itself.

Where is the official Jev SDK?

TypeSafe maintains its own Python and JavaScript clients, and that is where to start rather than any community project. A first working call is covered on the SDK page.

Which Jev GitHub project is production-ready?

On this evidence, none. No releases anywhere, half at five commits or fewer, and nothing measuring whether the decisions were right. jev-review is the best pattern to copy and fast-jev-compaction the most defensively built — a recommendation to read them, not to ship them.