browser-use/jev-ultrafast is first in this morning's tracked GitHub 7-day ranking, at 17,706 stars. The reported 7-day gain is 17,706.
That figure is the current total, not stars added across seven days. The first stored snapshot is 3,341 stars on September 18 at 08:01 UTC. The latest is 17,706 on September 22 at 16:00 UTC. The observed change is +14,365 stars over that shorter window, and the window opens after the repository already had thousands of stars.
Eight of the ten repositories at the top of this ranking show the same pattern: the reported 7-day gain equals the current star count.
Five movements worth separating
Production snapshot September 23, 2026, at 09:17 CEST (07:17 UTC). The public GitHub board's Last Updated was the same day, and the rows it rendered carried these star totals. Observation times differ by row. These are not one synchronized measurement.
Project | 7-day rank | Stars at latest observation | Board's reported 7-day figure | What the snapshots show | Latest observation (UTC) |
|---|---|---|---|---|---|
1 | 17,706 | 17,706 | 3,341 on Sep 18, 08:01 → 17,706 (+14,365) | Sep 22, 16:00 | |
2 | 233,321 | +8,703 | 224,618 on Sep 15, 08:20 → 233,321 (+8,703) | Sep 22, 18:20 | |
3 | 6,304 | 6,304 | 178 on Sep 18, 00:01 → 6,304 (+6,126) | Sep 23, 00:00 | |
4 | 6,287 | 6,287 | 4,297 on Sep 21, 08:00 → 6,287 (+1,990) | Sep 23, 00:00 | |
8 | 4,200 | 4,200 | 246 on Sep 19, 08:01 → 4,200 (+3,954) | Sep 23, 00:00 |
Source: GitHub board. Rankings cover the repositories Indie Signals tracks. The live page continues to change.
When the stored history is shorter than seven days, the reported 7-day figure is stretched toward a full week and then stopped at the current star count. For the four newer rows above, that stop is the number on the board. DeepSeek Harness is the exception in this table: +8,703 is the difference between the September 15 and September 22 snapshots. The +14,365 and the +8,703 cover different lengths of time, so they are not a race.
These are not one product
jev-ultrafast is Browser Use's browser agent. The README says TypeSafe's Jev picks an operation and an element, and a separate text model writes only when the operation is typing. Running the example requires a TypeSafe API key.
fast-jev-compaction is a Claude Code plugin that asks Jev which tool calls and tool results to keep. Text it keeps stays verbatim. The README's default model setting is jev-latest.
kev is a family of smaller decision models you train and run locally, with an API shaped like TypeSafe's. It is not the hosted Jev service.
ZCode is on this list because its reported 7-day gain equals its star total. The repository describes an AI coding workbench with a desktop app, a browser UI, and a terminal agent. It is not a Jev tool.
The same public 7-day list shows laya-mlx at 5,425 stars and a reported 7-day gain of 5,425. That README describes a local MLX runtime for Laya typed-decision models, and it says the port is independent rather than a TypeSafe or Convai release.
What the star count does not decide
TypeSafe's documentation describes Jev as a model that takes a state plus typed questions — choose an option, score a rubric, or say whether a statement is true — and returns structured values and probabilities rather than generated text. That shape fits a step your code already knows how to branch on. It does not say your users will accept the answers.
The jev-ultrafast README's timing note is narrower than the ranking: it describes repeated runs of one browser task and says that check is not a general reliability benchmark. Those seconds are the maintainer's measurement, not evidence of adoption.
A local model such as kev, and a plugin whose default is jev-latest, are different maintenance bets. Stars do not choose between them.
Ignore this inference: a repository at the top of the 7-day board, with a gain equal to its star count, has earned a place on your roadmap.
Verdict: WATCH
If one step in your product is a repeated closed decision — which element, which tool, keep or drop this result — read one of these implementations and try that decision on examples you already know the answer to. Keep the test small enough to throw away.
Do not add a hosted key or a local model because of the rank.
What would change the verdict? Move to BUILD when that test beats the chat-model call you use now on the errors you care about, you can pin the model or the checkpoint, and the upkeep is a cost you can carry. Move to IGNORE for your product if the step needs new prose, or the closed decision is wrong often enough that a person or a generative model still has to redo it.
What, if anything, changed in what you plan to build, buy, or ignore after reading this? Reply with one sentence.
P.S. Founding Topic Watch: stop tracking the AI ecosystem yourself. I’ll watch up to five Indie Signals topics or tracked projects that matter to your build / buy / ignore decision and tell you when the evidence changes — a Friday status email every week, plus an alert the moment the verdict moves. Eight weeks, five seats, $149 once. Reply WATCH and I’ll send the details.