Fresh Commits 05: Decision Models Went From One Repo to Eight in Nineteen Days

Fresh Commits ranks new AI repositories by how fast they’re gaining stars rather than by how many they’ve got, because a repo adding 700 a day tells you something a repo sitting on 90,000 since 2024 can’t. Every figure here came out of the GitHub API during this scan. This issue covers two weeks instead of one, since issue 04 published on September 20.

The headline: eight separate repositories built around decision models now exist, none of them older than nineteen days, holding 52,135 stars between them. Fastest is browser-use/jev-ultrafast at roughly 1,293 stars a day since launch. A second cluster turned up too, tools whose entire job is supervising the coding agents you already run. Z.ai’s ZCode carries a 30.5% fork ratio, nearly triple what a popular repo normally sees. Last week’s leader has added 2,461 stars without a single commit since September 18.

Best for anyone deciding what to actually install this month. Not ideal for anyone who wants finished software with documentation.

Three weeks ago, “decision model” wasn’t a category. It was one product with a weird pitch.

Now it’s eight repositories holding 52,135 stars, with a tooling layer forming around them in public.

That’s the thing worth paying attention to this fortnight. Not any single repo on the list below, though a few of them are very good. The pattern. When eight unrelated developers independently decide the same missing piece is worth building, the missing piece was real.

Here’s what the scan turned up. Here’s also what the numbers actually say once you stop reading star totals as if they mean something on their own.

How This List Gets Built

Every repo below was created inside the last 30 days. Stars, forks, licence and creation date were read from the GitHub API during this scan, not scraped off a trending page, so the figures are quotable rather than indicative.

One honest caveat on velocity. For a repo that’s four days old, “stars per day” is total stars divided by age, which is exactly what it sounds like and exactly as meaningful as it sounds. It’s a since-launch average. It isn’t today’s movement. Nothing here claims otherwise. Anywhere you see a per-day figure in this issue, that’s what it is.

Thirty-seven repos cleared the age bar this scan. Thirty of them carry a real licence, meaning 19% don’t. Most of the rest were skill packs, which is its own growth story and one we covered when the agent skills marketplace started filling up. That’s better than the rough third this column usually throws out, though “better than usual” is doing some work in that sentence.

Nine made the list.

The Ranking

Fork ratio is forks divided by stars. A star costs nothing. A fork means somebody pulled the code onto their machine with the intention of running it or changing it. Under 12% is normal for a repo people are mostly bookmarking. Above that, people are using the thing.

Keep an eye on row seven.

Eight Decision Models Repos In Nineteen Days

Before the individual entries, the pattern, because it’s the actual story.

Count the decision model repos that showed up in this scan under the 30 day bar. browser-use/jev-ultrafast at 21,973 stars. jaredpalmer/kev at 8,428. tamaratran/fast-jev-compaction at 7,375. mizorewww/laya-mlx at 6,754. TheoLeeCJ/SemIf-OpenJev at 4,685. dzhng/jevgrep at 2,243. lakeday-org/perch at 346. strands-labs/strands-decider at 331.

Eight repos. 52,135 stars. The oldest is nineteen days.

A decision model doesn’t write you an answer. It picks one or scores one, then hands back a number saying how sure it is. That’s the whole product. You ask it a yes or no question about a chunk of text or a screenshot or a diff. It answers in milliseconds rather than seconds, because it never generates prose on the way to the answer.

What’s happening in that list is a tooling layer assembling itself around that primitive. It’s happening fast enough that nobody has stopped to check the foundations. dzhng/jevgrep uses decisions to find files in a codebase. lakeday-org/perch uses them to lint code semantically. tamaratran/fast-jev-compaction uses them to decide which parts of a Claude Code conversation are still worth keeping. strands-labs/strands-decider is a decision model that reports a calibrated confidence on every call.

Nobody Has Checked The Number Everything Rests On

That last description is the one to sit with. Calibrated confidence means the number is supposed to be trustworthy: when it says 70% it should be right roughly 70% of the time. Almost every project in this category advertises a confidence score. Almost none of them publish the data showing the score is accurate. Eight repos now depend on that number being meaningful. The burden of proof drifts further behind the build rate every week.

That’s not a reason to avoid them. It’s a reason to not wire one into something that matters without testing it on your own data first.

1. browser-use/jev-ultrafast

Seventeen days old, 21,973 stars, MIT, Python. Last commit September 30.

Billed as the fastest and cheapest web agent. It comes from browser-use, which already had an audience before this repo existed, so some of that star velocity is a mailing list rather than a discovery. Worth knowing when you compare it against row three, where nobody had heard of the author a week ago.

The fork ratio is 7.1%, which is the lowest interesting number on the whole table. Twenty two thousand people starred it. Fifteen hundred pulled it down. For a tool whose entire pitch is speed and cost, the gap between people who liked the idea and people who ran the benchmark is wide.

Speed claims on web agents are also the easiest claims in this field to make and the hardest to compare, because nobody agrees on what a page load costs or which sites count. If you’re evaluating this, measure cost per completed task on your own sites. Latency on a demo tells you nothing you can budget against.

Still the fastest climb on the list by a clear margin. If you’re doing browser automation, this is the one to look at first.

2. Niko1221/Strata

Ten days old, 10,283 stars, MIT, C++. Pushed today.

One click local inference for Qwen3.8-Flash-Next on consumer hardware, Windows or Linux, serving an OpenAI and Anthropic compatible API on localhost with optional image input.

This is the most useful repo on the list for a normal person, which is not the same thing as the most impressive. Ten thousand stars in ten days for a one click installer tells you how much appetite there is for running a capable model locally without a weekend of CUDA archaeology. The API compatibility is the real feature: point anything that already speaks to OpenAI at localhost instead, change nothing else.

The 8.8% fork ratio is slightly low for an installer, though installers are the one category where that’s expected. You don’t fork a one click installer. You run it.

C++ with 909 forks and daily commits is a project with actual contributors rather than a weekend release that went viral. That matters more for whether it still works in March than any number on this table.

3. rehan-remade/universal-modder

Four days old, 2,962 stars, MIT, Python. Pushed today.

Point Claude at any PC game you own. It bundles skills, tools and an MCP for recon, reverse engineering, generated art and audio, in game testing plus showcase videos.

Seven hundred forty stars a day from a standing start is the most impressive velocity on the list once you adjust for audience, because this author didn’t have one. It also arrived the same week an X post about AI game modding was circulating, which is how a repo like this goes from zero to three thousand in four days.

Treat the capability list with the skepticism a four day old repo deserves. “Reverse engineering” plus “recreate features for your own stack” is a category where what works on one game engine fails completely on the next. Nothing in a 96 hour old project has been tested across enough games to know which. The modding community will find the edges within a fortnight.

Also worth saying plainly: modding games you own is fine. The licensing picture for anything you distribute afterwards is not something an MCP server resolves for you.

4. feder-cr/dots

Four days old, 2,597 stars, MIT, Python. Last commit October 3.

An open source agent with its own browser, pitched specifically as one that doesn’t get blocked.

A 16.8% fork ratio on a four day old repo is the second strongest signal on this table. Four hundred thirty seven people cloned this inside 96 hours. That’s not bookmarking behaviour. People have a specific site they want to automate and they’re trying it right now.

The anti blocking claim is the one to watch and the one most likely to quietly stop being true. Bot detection is adversarial. A technique that works the week it ships is a technique the detection vendors have a sample of by the end of the month. Nothing wrong with that, it’s the nature of the problem, but it means the repo’s value decays in a way a normal tool’s doesn’t. On this one, a three month old commit history would be a warning rather than a reassurance.

Name collision warning, which is this issue’s small annoyance: this is not CopilotKit/OpenDots below. Neither is the “dots” being discussed in agent threads on X right now. Three different things, similar names, one week.

5. CopilotKit/OpenDots

Five days old, 3,100 stars, MIT, TypeScript. Last commit October 2.

Always on AI coworkers that move between text, calls and Slack.

CopilotKit is an established project, so this lands with an audience the way jev-ultrafast did. The 13.1% fork ratio is healthy for a five day old repo in a category where most people want a hosted thing rather than a cloned thing.

What it signals matters more than what it does. Something that lives in Slack and answers calls is not a coding tool, it’s a colleague shaped product, which is a different bet about where this goes than most of the list is making. If you’ve read our breakdown of Claude’s computer use and Dispatch setup, this is the same wager from the open source side: the value sits in being always available wherever work already happens, not in the model.

Five days is far too early to judge whether it holds up under a real Slack workspace. Star it, check back in a month.

6. Louis-CFM/coucou

Six days old, 3,396 stars, MIT, Swift. Pushed today.

A small creature that lives in your Mac’s notch, or at the top of the screen on Windows and Linux, watching your coding agents: Claude Code, Codex, Cursor, Gemini CLI, Antigravity and others.

This is the one I’d actually install. It also points at the second pattern in this scan.

Nobody needed a notch pet in 2024. Nobody was running four agents at once. The 15.4% fork ratio on a Swift menu bar app is unusually high. The reason gets obvious the moment you’ve left Claude Code waiting on a permission prompt for twenty minutes while you did something else. It doesn’t solve a technical problem. It’s that agents are now slow enough and numerous enough to need supervision. Watching four terminals is not supervision. It’s just four terminals.

Look at the rest of the scan and the cluster is unmistakable. coucou watches your agents. yetone/magpie at row nine routes them between models. OpenDots gives you agents that report in. useagenthq/useagent, 35 days old so just outside the bar, hands back finished work from agents running on their own cloud machine.

Four separate projects whose product is agent management rather than agent capability. That’s new this month. It’s the same pressure that pushed people to run Claude Code out of Telegram and Discord, just solved from the desktop instead. If you’re already running several at once, our rundown of GitHub repos that make Claude Code better covers the rest of that stack.

7. zai-org/ZCode

Fourteen days old, 7,401 stars, Apache-2.0, TypeScript. Last commit September 29.

Z.ai’s own coding agent. The description says powerful, intelligent and extensible, which tells you nothing, so look at the forks instead.

Two thousand two hundred fifty nine forks on 7,401 stars. That’s a 30.5% fork ratio.

Nothing else on this table is close. Normal is under 12%. The second highest here is 16.8%. ZCode is nearly double that and almost triple the baseline, which means one of two things is true. Either an unusual share of the people who found it immediately started running and modifying it, which for a coding agent would be a very strong signal, or it’s structured as something you’re meant to fork to use, in which case the ratio is an artefact of the design rather than evidence of enthusiasm.

I can’t tell you which from the API. You can tell in about five minutes by reading the install instructions. Those five minutes are worth spending, because if it’s the first one then this is the most under discussed repo in this issue by a distance. A lab shipping its own coding agent, Apache licensed, with a third of its audience running the code.

The September 29 last commit is the only thing giving me pause. Five days of quiet on a two week old project is nothing. Five more would be something.

8. mizorewww/laya-mlx

Fifteen days old, 6,754 stars, Apache-2.0, Python. Last commit October 2.

A native MLX runtime for Laya typed decision models, reporting 7 to 14 milliseconds for short decisions on an M3 Max, with no text generation, no PyTorch and no cloud API.

Those latency figures are the developer’s own and no independent benchmark exists, so treat them as a claim rather than a measurement. The shape of the claim is plausible though. Skipping text generation is exactly why these models are fast, plus MLX on Apple silicon is a reasonable place to see numbers like that.

What makes this one matter is the “no cloud API” part. Every decision model pitch so far has been about cost per million calls, which assumes the calls leave your machine. A local runtime changes the calculus completely: the per decision cost goes to zero and the question becomes whether your laptop is free. For anything running a decision on every keystroke or every file in a repo, that’s the difference between a feature and a budget line.

Eight percent fork ratio, which for a runtime aimed at Apple silicon owners is about right.

9. yetone/magpie

Eleven days old, 4,785 stars, MIT, Go. Pushed today.

Every agent’s model in one place, from the menu bar. Codex running on DeepSeek, Claude Code running on Kimi.

The pitch is model routing without touching config files. It exists because the economics of running coding agents got strange. Paying for two or three agent subscriptions while also holding API credits for open models is now a normal setup. Nothing was managing it. Our Claude Pro review walks through where the subscription actually makes sense versus paying per token, which is the decision magpie is trying to make reversible rather than permanent.

The 7% fork ratio is the low end of this table. For a menu bar utility written in Go that’s expected. People download binaries. They don’t fork them.

Daily commits on an eleven day old project from an author with a track record. Reasonable bet.

The Repo Gaining Stars With Nobody At The Wheel

tamaratran/fast-jev-compaction led issue 04 at 4,914 stars. Today it’s at 7,375.

Its last commit was September 18.

So it has added roughly 2,461 stars across two weeks in which nobody touched the code. That’s the cleanest example this column has run into of stars measuring attention rather than activity. It’s why the fork ratio column exists in the table above.

To be fair to it, the idea is good. Replacing Claude Code’s compaction summary with scored keep or drop decisions, so that what survives is verbatim rather than paraphrased, is a clever fix for a real annoyance. A tool solving one specific problem can be finished. Not every repo needs weekly commits to be worth installing.

But it’s a 17 day old project in a category changing weekly, built against a plugin interface that is itself changing. The author has been elsewhere for the last 16 days. Those two facts sit uncomfortably together. If you install it, pin the version and expect to understand it yourself when something shifts underneath it.

Last Week’s Picks, Two Weeks Later

Issue 04 published on September 20, which makes this a 14 day follow up rather than a 7 day one.

Three of the seven appear in this scan with verified figures. tamaratran/fast-jev-compaction went from 4,914 to 7,375, covered above. jaredpalmer/kev went from 820 to 8,428, which is the largest move any pick this column has made and roughly a tenfold increase in a fortnight. laya-mlx was not an issue 04 pick, so despite the family resemblance it counts as new this week.

kev deserves a second look because of what it is. A half billion parameter model for typed decisions that trains and runs on a MacBook, Apache licensed. At 820 stars it read as a curiosity. At 8,428, with 553 forks, it reads as the thing that made the rest of this week’s list possible. If you want one entry point into the decision model category, start there rather than with the fastest climber.

The remaining four are jarrodwatts/jev-trader, awlevin/typesafe-computer-use, korcarc/text-humanizer plus saragordic/window-sweaters. None of them are in this scan, which proves nothing. The scanner suppressed 113 entries as already seen this run, so a pick being absent means it wasn’t re surfaced rather than that it stopped moving. Reporting a repo as dead on a cache miss is exactly the kind of mistake this column shouldn’t make. Those four links go straight to the source so you can check the counts yourself. Verified figures follow in an update rather than a guess.

A Correction On Last Week’s Rates

One more correction worth making in public. Issue 04’s per day figures were produced by a version of the scanner with a known velocity bug, which reported weekly star gains as daily ones. Where last week’s rate and this week’s total disagree, this week’s API read is the one to trust. The totals in 04 were fine. Some of the rates were not.

What You Should Actually Do

Install two things, ignore the rest for a month.

If you run more than one coding agent at a time, install coucou. It’s the lowest risk item on this list. It solves a problem you definitely have. A Swift menu bar app also has roughly no blast radius if it turns out to be bad.

If you want to run a capable model on your own hardware, install Strata. One click, API compatible with what you already use, ten thousand people got there before you this week.

Everything else on the list is worth a star and nothing more yet. Four and five day old repos are not software, they’re intentions with a README. The ones that survive October will still be there in November and will have docs by then.

On the decision model cluster specifically, there is one rule. Before you wire a confidence score into anything that makes a decision you’d care about getting wrong, take 100 examples from your own data where you already know the right answer, run them through, then check whether the number means what it says. If a project advertises a calibrated confidence and can’t show you the calibration, you’re the one calibrating it.

That takes an afternoon. Finding out in production takes longer.

/separator

The Part Worth Keeping

Eight repos, 52,135 stars, nineteen days, one idea.

The speed is the finding. Fast is not a virtue by itself. What matters is that nothing in that list has existed long enough for anyone to have checked whether the thing underneath it works as advertised. A confidence score is a promise about how often it’s wrong. Eight projects are now built on that promise. The number of them publishing evidence for it is still zero.

The category is real. Tooling is arriving faster than verification, which is the normal order of events and worth naming anyway.

Check back next week. Something on this list will be gone.

Charts and Blocks

Stars Per Day Since Launch

The Decision Model Cluster