Fresh Commits runs every Sunday. It ranks new AI agent repositories by stars gained per day rather than total stars. Every figure comes from the GitHub API, then we check it before publication. That includes the licence, since roughly a third of what surfaces each week is not open source. This week: seven repos cleared verification and five of them are built on Jev, a model TypeSafe released on Monday. The top entry took 4,914 stars in three days. One repo holds a 19% fork ratio on day three, the highest this column has recorded. In the revisit, last week’s lead pick now trips our own bought-stars check, one featured repo has vanished from GitHub entirely, plus another has gone quiet for ten days. Best for anyone tracking where agent tooling is heading. Not ideal as a list of things to install today.
Seven repos cleared verification this week. Five of them are built on a model that did not exist on Sunday.
TypeSafe released Jev on Monday. By Friday it accounted for most of this column.
That has not happened before. Usually a week’s list spreads across unrelated problems, meaning somebody’s Blender pipeline sitting next to somebody’s usage tracker.
Also this week: a repo from last week’s issue has disappeared from GitHub. The repo we led with seven days ago now fails the check we run on everything.
How This Week Ranked
Sorted by stars per day. Every row verified through the GitHub API on September 20.
| Repo | Stars | Per day | Licence | Forks | Language |
|---|---|---|---|---|---|
| tamaratran/fast-jev-compaction | 4,914 | 1,638 | MIT | 272 | TypeScript |
| TheoLeeCJ/SemIf | 2,308 | 577 | MIT | 143 | Python |
| jarrodwatts/jev-trader | 1,463 | 488 | MIT | 273 | TypeScript |
| jaredpalmer/kev | 820 | 410 | Apache 2.0 | 53 | Python |
| awlevin/typesafe-computer-use | 582 | 194 | MIT | 35 | Python |
| korcarc/text-humanizer | 734 | 184 | MIT | 84 | Python |
| saragordic/window-sweaters | 614 | 123 | GPL 3.0 | 18 | C |
1. fast-jev-compaction, 1,638 Stars a Day
fast-jev-compaction is a Claude Code plugin replacing the compaction summary with Jev decisions on every tool call plus result.
What it does
When a Claude Code session fills its context window, compaction summarises what came before so work can continue. The model writes that summary, which costs tokens plus loses detail nobody controls.
This swaps that for typed decisions from Jev on each tool call, which runs faster plus cheaper because Jev returns structured values rather than prose.
Why the numbers stand out
4,914 stars in three days is the fastest climb this column has recorded. 272 forks behind it, 56 open issues, MIT licensed.
Fifty-six open issues on a three-day-old repo reads as heavy real usage rather than neglect. People installed this into working setups then immediately found edges.
The one concern
Last push was two days ago, which on a repo moving this fast is worth watching rather than worrying about. Check the commit history before you install it into anything that matters.
2. SemIf Runs the Same Idea on Your Own GPU
SemIf took 2,308 stars over four days. MIT, Python, 143 forks, homepage at openjev.com.
Semantic ifs on a 3090
The pitch is semantic conditionals from open models running on a 3090 at home. Rather than calling a hosted decision endpoint, you run the decision locally on consumer hardware.
The repo states plainly that it is independent and not affiliated with Jev or TypeSafe, which is more disclosure than most compatible projects bother with.
Why local matters here
Jev is gated behind a waitlist plus priced per million input tokens. A local alternative removes both constraints, so you trade absolute quality for control plus predictable cost.
Nobody has answered whether it matches Jev on calibration, which is the question nobody has tested on any of these. Eleven open issues suggest people are finding out the hard way.
3. jev-trader Has the Highest Fork Ratio We Have Recorded
jev-trader took 1,463 stars over three days with 273 forks, MIT, TypeScript.
The ratio
273 divided by 1,463 is 18.7%. Roughly one in five people who starred it also cloned it.
We started tracking fork ratios two weeks ago because stars cost a click while forks cost intent. Nothing has come close to this. Last week’s leader sat at 32% on day one with far smaller absolute numbers, so this is the highest ratio at meaningful scale.
What people are cloning
One AI trade decision every Monad block, running Jev against a MON-USDC pair on Kuru.
The caution
People are forking a trading bot built on a model whose calibration nobody outside the vendor has tested. TypeSafe itself advises testing confidence thresholds against your own labelled examples before relying on them.
Three open issues against 273 forks is a low ratio, which usually means people cloned it to read rather than to run. Worth hoping that holds.
4. kev Trains on a MacBook
kev took 820 stars over two days. Apache 2.0, Python, pushed today.
The approach
A tiny Jev-like model built on Qwen2.5-0.5B that you can train plus run on a MacBook.
Half a billion parameters with the training loop included. The interesting test embedded here is how much model a typed-decision task actually needs, since if 0.5B handles most routing then the category is cheaper than anyone assumed.
Why Apache 2.0 reads as deliberate
Apache carries an explicit patent grant that MIT leaves out. For a project reimplementing a commercial product’s approach, that choice suggests somebody thought about it.
Pushed today, 53 forks, 4 open issues. The most actively maintained repo on this week’s list.
Stars, forks plus open issues on all seven
Licence identifiers, including five rejections
Creation plus last push dates
Last week’s picks rechecked for the revisit
One repo confirmed missing from search
Any of these matching Jev’s calibration
SemIf running usably on a 3090
kev training on a MacBook
Any humanizer defeating detection
Every TypeSafe performance figure
Figures pulled September 20, 2026. Nothing here was installed or run. Star counts move daily.
5. typesafe-computer-use, Cheap by the Step
typesafe-computer-use took 582 stars over three days. MIT, Python, pushed today.
OCR the screen, classify the next action with a typed decision, click. Around $0.0002 a step.
Why the price is the pitch
Computer use normally means a frontier model reading a screenshot then reasoning about what to click, which costs real money per step plus takes seconds.
Two hundredths of a cent per step puts computer use somewhere it has not been, meaning loops that run continuously rather than on demand.
What to check before running it
Eight open issues against 35 forks. Low fork count relative to its stars, so most people are reading rather than running.
Read anything through before letting it OCR your screen then click on it. This one is small enough that reading it is realistic.
6. Another text-humanizer, in a Category That Got Crowded
korcarc/text-humanizer took 734 stars over four days. MIT, Python, 84 forks.
The same pitch as last week
Converts AI-generated text into something reading as human written. We covered a different repo with the same name plus the same claim in last week’s issue.
How big this category actually is
A search for text-humanizer now returns 867 repositories. Two of them dwarf everything: one at 5,650 stars plus another at 3,042, both MIT, both around four to five months old.
Neither qualifies for this column, since both fail the under-30-days rule. Their existence reframes this week’s entry though, because 734 stars makes it third in a category with two established leaders.
The honest assessment
Tools promising to defeat AI detection have a poor record. Most swap vocabulary plus vary sentence length, addressing surface tells while leaving structural ones intact.
The demand is real regardless. Three repos with the same premise trending inside five months tells you plenty about how many people are publishing AI-assisted text at volume.
7. window-sweaters, Because Not Everything Is Agents
window-sweaters took 614 stars over five days. GPL 3.0, C, 18 forks.
A macOS menu-bar app dressing your windows in knitted borders, so each app gets its own colour.
Why it is here
It solves a genuine problem, which is telling windows apart at a glance, in a way nobody would have specified in a requirements document.
GPL 3.0 rather than MIT, meaning derivative works must stay open. An unusual choice for a small utility plus a deliberate one.
Eighteen forks against 614 stars is a 3% ratio, the lowest on this week’s list. People like it more than they modify it, which is exactly right for a finished small tool.
Why One Model Took Over a Whole Week
Five of seven is unusual enough to explain rather than just report.
The interface was published, the weights were not
TypeSafe documented how Jev behaves. Typed questions in, typed answers out, confidence values attached, a documented request budget.
That is enough to build against without access, plus enough to reimplement if you cannot get access. Both routes appear on this week’s list, since fast-jev-compaction and jev-trader build on the hosted model while SemIf and kev rebuild it.
Access was gated
Jev sits behind an early access waitlist. Anyone impatient with a GPU had an obvious alternative, which is why two of this week’s seven are local reimplementations rather than integrations.
Gating a launch is reasonable. It reliably produces open alternatives, which is roughly what happened to every gated model release of the past three years.
The problem was already annoying people
Everyone shipping agents has been writing structured-output plumbing by hand. Parsers, validators, retry loops, guardrails around a model that occasionally invents a category.
A format landing on an existing pile of frustration spreads faster than one creating a new category. We saw the same shape when agent skills spread, plus when harnesses became portable.
What this predicts
Concentration like this does not last. Either the clones converge on something good enough that Jev becomes one option among several, or somebody publishes a calibration comparison that settles whether the hosted version is meaningfully better.
Either way next week’s list looks different. Weeks where one thing owns the column are usually the peak rather than the trend.
The Fork Ratio, Ranked
Stars are cheap. Forks cost intent, so the ratio between them says more than either number alone.
This week’s spread
| Repo | Fork ratio |
|---|---|
| jev-trader | 18.7% |
| text-humanizer | 11.4% |
| kev | 6.5% |
| SemIf | 6.2% |
| typesafe-computer-use | 6.0% |
| fast-jev-compaction | 5.5% |
| window-sweaters | 2.9% |
Reading it
jev-trader at 18.7% is the highest we have recorded at this scale. A trading bot draws that because people want to run it against their own pair rather than read about somebody else’s.
fast-jev-compaction sits at 5.5% despite having six times the forks in absolute terms, because its star count is enormous. A plugin gets installed rather than cloned, so a low ratio there means nothing bad.
window-sweaters at 2.9% is exactly right for a finished small tool. Nobody forks a menu-bar utility they like, they just use it.
The limit of this metric
Ratios only mean something once a repo clears a hundred stars or so, plus they measure different things for different project types. A library, a plugin and a template should not be compared on the same axis.
Worth tracking anyway, since it is the only signal available that costs the person more than a click.
What Got Cut
Five repos failed verification. Three outrank most of what passed.
The licence failures
mcncarl/jianying-headless sits at 1,874 stars with no licence GitHub can classify, describing itself as a private source preview.
bespokelabsai/nimble has reached 959 stars. Bespoke Labs announced it nine days ago as an open data, open model, open recipe project. There is still no licence file, which we wrote about on Friday when it stood at 33 stars. Twenty-nine times the stars since then, no change to the terms.
yibie/awesome-jev at 550 stars plus yynxxxxx/gpt_sub_analysis at 566 both lack licences too. The second describes itself as an internal draft under coordinated disclosure, so it stays out on those grounds as well.
The bought-stars signal
mizorewww/laya-mlx at 1,173 stars, Apache 2.0, a native MLX runtime for typed decision models claiming 7 to 14 millisecond decisions on an M3 Max.
Zero open issues at that star count trips our check. A repo past 500 stars with nobody filing anything usually means the stars did not come from users.
This one looks legitimate otherwise, so it gets rechecked next week rather than dismissed. The rule exists because it catches things, plus it produces false positives.
Last Week Revisited
Every issue reports back. This week the revisit is more interesting than some of the new entries.
One repo has disappeared
SpaceDudem/text-humanizer featured in last week’s issue at 727 stars. It now returns zero results on GitHub search, so there is deliberately no link here. There is nothing left to link to.
Deleted, renamed or made private. No way to tell which from outside. A repo we verified seven days ago at 727 stars is simply gone, which is worth remembering the next time a roundup tells you something is worth building on.
Our lead pick now fails our own check
kruzovic7/ai-data-extractor led last week at 296 stars with a 32% fork ratio, which we called the highest intent-to-run we had recorded.
It has grown to 842 stars. It also still has zero open issues, so at 842 it now trips the bought-stars rule that only applies above 500. Last week it sat under the threshold.
Which means our own lead pick would be rejected this week. The check does not know it recommended the repo seven days ago, which is rather the point of having one.
codenotch is holding
codenotch sat at 1,532 stars last week, 2,112 now. Forks up from 240 to 328. Pushed today, sixteen consecutive days of commits.
Twenty-two open issues, which is healthy for a tool this widely installed. The strongest survivor from previous issues.
dream-loop has gone quiet
dream-loop had 953 stars last week, 1,461 now, so attention kept climbing. Last push was ten days ago.
We flagged it at three days stale last week and said we were watching. Ten days is not abandoned, though it is four days from our fourteen-day cutoff. If nothing lands this week it fails the next issue.
What This Column Has Got Wrong So Far
Four issues in, the revisit has caught enough mistakes to be worth listing plainly.
The vanishing repo
SpaceDudem/text-humanizer passed every check on September 13. Real licence, active pushes, reasonable issue count, 727 stars. Seven days later it does not exist.
No check catches that. Nothing in the API on the day of publication indicated a repo about to disappear, which means the roundup format has a floor on how reliable it can be.
The threshold problem
ai-data-extractor passed last week at 296 stars with zero issues, then fails this week at 842 with zero issues. Our rule only applies above 500.
Nothing changed about the repo except its popularity. Either that threshold should sit lower, or zero issues means something different on a two-day-old project than on a two-week-old one. Probably the second.
The false positive
mizorewww/laya-mlx got cut this week for zero issues at 1,173 stars. Everything else about it looks legitimate, including an Apache 2.0 licence plus a specific technical claim about MLX runtime latency.
The rule catches bought stars. It also catches well-built things nobody has broken yet, with no way to tell those apart from outside.
Why publish the misses
A column that only reports its hits is a marketing channel. The checks here reject roughly a third of what trends, which only means something if the failures get reported alongside the passes.
What You Should Actually Do
Check the licence before the stars
Five repos got cut this week and four of them had no usable licence. One was announced as open nine days ago plus has since grown to 959 stars without adding one.
It takes ten seconds to look. It decides whether you can legally build on anything you clone.
Read the fork ratio
Stars measure attention, forks measure intent. This week ran from 18.7% on jev-trader down to 2.9% on window-sweaters.
Neither is better. A high ratio means people are running it. A low one means people admire it. They tell you different things about whether anyone has tested the thing besides its author.
Watch the push date on fast movers
A repo taking 1,638 stars a day is changing underneath you. Check the last commit before you install anything from this list into work that matters.
Do not assume a repo will still exist
That is the lesson from last week’s vanished entry. If you depend on something small, fork it yourself.
The Part Worth Keeping
Five of seven entries built on a model that launched six days ago. A vanished repo. A lead pick that now fails its own gate.
The Jev concentration is the headline, though the revisit matters more. Every week this column publishes a list of things climbing fast. The honest follow-up is that one of them disappeared inside seven days while another accumulated 546 stars without anybody filing a single issue.
Fast-moving does not mean durable. It barely means real.
Check the licence, read the fork ratio, then look at when somebody last touched it. Three checks, about thirty seconds. They would have caught everything this column got wrong so far.
Charts and Blocks
This week ranked by velocity
Last week’s picks, seven days on
