The Seven: Agent Repos Worth Your Saturday (Issue 01)

Fresh Commits is a weekly GitHub roundup covering AI agent tooling, ranked by stars gained per day rather than by total stars. Every figure here came out of the GitHub API and passed a verification check before publication, so each entry prints its exact star count, age, plus license for you to check yourself. Three projects launched within the last 48 hours, including an evolutionary agent runtime that was one day old at press time. Four are older projects with real momentum and almost no English coverage, labelled by age instead of passed off as new. From issue two onward, every Fresh Commits closes by revisiting the previous week’s picks to report which kept climbing and which stalled. Three repos got cut during verification and the reasons are printed too. Best for developers who want to find tools early. Not ideal for anyone wanting finished, polished software.

Most GitHub roundups rank by total stars. That returns the same projects every week, because a 40,000 star project will still have 40,000 stars next Saturday.

This one ranks by stars gained per day.

A repo that picked up 145 stars yesterday scores 145. A project sitting on 40,000 stars since 2024 scores about six. The second is more popular. The first is what’s actually moving.

Every number below came from the GitHub API on August 8 and went through a verification pass before anything got written. If a figure looks different when you click through, the repo moved after publication, which happens fast at this end of the scale.

How These Get Picked

Worth stating plainly, since a roundup without criteria is a list of whatever the writer noticed.

Candidates come out of the GitHub API filtered to AI, agent, plus developer tooling, then ranked on stars per day since creation. Each one then goes through a check covering age, license, staleness, plus issue-to-star ratio.

That check split this issue in two. Three repos are days old. Four have been around for months but carry momentum and no English-language coverage. Both groups are worth your time. Only one group is new, so they sit in separate sections with ages printed instead of blurred together.

There’s a ceiling at 20,000 stars. OpenClaw gained roughly 1,500 stars a day this week and Hermes Agent gained 597, which is real. But those sit at 385,000 and 227,000 stars, so leading with them every Saturday would produce a column nobody needs.

License gets checked too. A repo with no license file is not open source whatever the README says. Installing one means you have no legal right to use the code.

New This Week

1. dgx-spark-2-deepseek-flash-0731

145 stars · 145 a day · 1 day old · MIT · Shell maliubiao/dgx-spark-2-deepseek-flash-0731

Highest velocity on the list. It isn’t software either. It’s a setup guide for running DeepSeek Flash across two DGX Spark boxes from scratch.

That says something about where demand actually sits. A shell script plus documentation for one specific model on one specific pair of machines outpaced every framework this week. Eighteen forks inside a day means people are adapting it rather than bookmarking it.

Documentation repos are chronically undervalued. The hard part of local inference was never the model, it’s the four hours of driver versions and memory configuration between you and something that runs. Somebody did those four hours then wrote them down.

Zero open issues on a one day old repo tells you nothing yet. Check back next week.

2. KADATH

133 stars · 133 a day · 1 day old · Apache 2.0 · Python i3T4AN/KADATH

Created August 8. Hours old when this went to press, which is exactly the window this roundup exists to catch.

An evolutionary multi-agent runtime that breeds, evaluates, then improves autonomous agents across reproducible epochs. Agents get scored, better ones produce variants, worse ones die out.

Put it beside Prime Agent, which we covered this week. A pattern appears. Prime Intellect’s harness rewrites its own prompts and skills mid-session. KADATH runs selection across generations instead. Both bet the interesting work has moved from the model to the scaffolding around it.

Reproducible epochs is the phrase to notice. Plenty of evolutionary agent projects produce results nobody can replicate, so building determinism in from day one suggests somebody thought before writing code.

One fork, zero issues. Nobody has tried hard enough to break it yet.

3. shuohao-skills

246 stars · 123 a day · 2 days old · Apache 2.0 · JavaScript eternityspring/shuohao-skills

A skills collection for coding agents that runs on both Claude Code and Codex.

Growth is coming almost entirely from Chinese-speaking developers. English-language AI coverage has nothing on it. That blind spot deserves naming. A meaningful share of the most active agent tooling right now ships with Chinese-language READMEs, so reading only English roundups means seeing maybe half of what exists.

The skills lean creative rather than engineering. One breaks a novel into character sheets with portraits, illustration prompts, voice prompts. Not what most people mean by agent tooling, which is part of why it’s interesting.

Twenty-seven forks in two days. Pairs with our GitHub repos roundup for anyone building a skills library.

4. opendot

26 stars · 2.4 a day · 11 days old · MIT · Python vedaant00/opendot

A terminal AI agent you can fully undo. It snapshots every file operation plus every shell action first, so you can see exactly what it did then walk it back cleanly.

Certainly the most quietly useful idea here.

Everyone building agent tooling optimises for capability. This one optimises for the moment your agent does something enthusiastic and wrong at 11pm. The maintainer’s framing: an agent you can let loose because nothing it does is a surprise, with almost nothing irreversible.

That “almost” is where the design gets good. Commands whose effects escape your workspace, meaning network calls, sudo, git push, or deleting outside the working directory, trigger a confirmation before running, with an honest note about what cannot be reversed. External tool calls through MCP work the same way, confirming first then marking the call irreversible in the ledger. Most agent tools pretend that distinction does not exist.

Model-agnostic through LiteLLM, so OpenAI, Anthropic, Google, or DeepSeek all work. It also runs fully local via Ollama. Drop an OPENDOT.md in a project and its contents become agent context.

uvx opendot
uv tool install opendot

Thirty forks against 26 stars is a strange ratio. More people cloned it than starred it, which usually means the people who found it found it because they needed it.

Moving, but Not New

Four projects with months behind them. Each has momentum plus almost no English coverage, which counts as undiscovered even when it isn’t recent.

5. smux

1,499 stars · 9.3 a day · 161 days old · MIT ShawnPana/smux

A tmux configuration with terminal automation plus agent-to-agent communication built in.

The human-facing half is ordinary: keyboard-driven config, Option-key bindings, pane labels. The other half is not. A CLI called tmux-bridge lets any agent read, type into, or send keys to any pane.

Which means Claude Code in one pane can prompt Codex in the next pane. Codex answers back.

tmux-bridge read codex 20
tmux-bridge type codex "review src/auth.ts"
tmux-bridge keys codex Enter

The protocol enforces read, then act, then read. You cannot type into a pane you have not read first, which stops agents talking over each other.

Anything that can run bash can join in, so the compatibility list runs past 40 agents. It’s a hack in the good sense. No protocol, no daemon, no orchestration framework. Just panes and keystrokes.

Its author, @shawn_pana, works on browser-use, the 107,000 star browser automation project. He announced smux in March. Japanese plus Chinese developers have published test threads since. English coverage: essentially none.

Worth reading beside the AISI report from this week, where agents left instructions on GitHub for other agents and later runs found them. Deliberate agent-to-agent messaging is arriving from several directions at once. This is the version you can install today.

6. Ghost in the Droid

298 stars · 2.4 a day · 122 days old · MIT ghost-in-the-droid/android-agent

Give any LLM agent a real phone.

Android runs over ADB, iPhone over WebDriverAgent with no jailbreak. It exposes 62 MCP tools covering tap, swipe, type, screenshots, accessibility tree reading, app launch, camera, text to speech. Nothing gets installed on the device.

The number worth checking: driven by Claude Code, it completes 115 of 116 tasks on AndroidWorld, Google Research’s Android agent benchmark, on the unmodified upstream harness. That’s 99.1%. The project labels this an early result with the methodology writeup pending, so treat it as a preview until the trajectories publish.

Platform, brain, driver are all independent. Point Claude Code at it, or Codex, or Cursor, or a model running on the phone itself through llama.cpp or MLX. On-device inference means it works in airplane mode with no cost per action.

It’s the only entry here with real media. There’s a demo on YouTube, a site carrying nine demo reels, plus an active account at @ghostinthedroid.

uvx --from ghost-in-the-droid android-agent-mcp

Anyone who read our OpenClaw deep dive will recognise the shape. Same idea, different body.

7. hal0

63 stars · 0.7 a day · 85 days old · Apache 2.0 · Python Hal0ai/hal0

Self-hosted home inference built for AMD Strix Halo, with multi-backend slots plus an OpenAI-compatible gateway.

Generally, local inference tooling assumes NVIDIA. Strix Halo owners get told to figure it out. This is somebody figuring it out then publishing, which is the least glamorous and most useful category of open source there is.

One /v1/* API covers chat, completions, embeddings, reranking, speech to text, text to speech, plus image generation, with the box scheduling concurrent workloads. It targets Ryzen AI Max chips but falls back to NVIDIA RTX 30/40/50, Radeon RX 7000, or plain CPU. Unified memory addressing reaches 124 GiB, which is the actual reason people buy this hardware.

One detail that matters here specifically: hal0 ships Hermes as a packaged extension, bootstrapping itself sandboxed under its own user, prewired to the local API plus your MCP servers, with tool-approval gating. Anyone who read our Hermes Agent coverage gets a one-command path to running it locally.

Seventy-six open issues against 63 stars is the figure to weigh. That ratio means either the maintainer documents everything they hit or the project is fragile. The maintainer is not accepting outside pull requests yet either, so treat it as one person’s project with an audience.

8. human

61 stars · 0.4 a day · 165 days old · MIT · Go gethuman-sh/human

Tickets, docs, designs, analytics go in. Shipped code comes out. They call it an AI software factory.

The ambition is enormous and probably too large. What makes it worth watching is the implementation choice: written in Go as a single pipeline rather than assembled from Python glue and prayers. Four open issues, live site at gethuman.sh.

Every part of this exists separately already. Doing all of it in one pipe is either the obvious consolidation or a project that collapses under its own scope. Ask again in a month, which is what the revisit section exists for.

Lowest velocity here at 0.4 stars a day, included because the idea is the most ambitious on the list and because a Go implementation suggests somebody serious about it running unattended.

What Got Cut

Three repos scored well on velocity but didn’t make it, worth showing since a list with no rejections isn’t curated.

MIgHTy-alIeN/ai-trader-bot gained 118 stars a day. It’s a crypto arbitrage bot with no license file, a description reading like SEO filler, plus a star pattern that looks purchased. Unlicensed plus crypto plus suspicious growth is an automatic cut.

3169657175/gpt-webcodex at 97 a day lets the ChatGPT web interface read and modify local files. Interesting idea, MIT licensed, but the whole approach depends on a browser interface OpenAI can change without notice. Shelf life measured in weeks.

GOROman/vibewatch at 95 a day is an M5Stack hardware stopwatch controller for AI-assisted coding. Charming, MIT, niche in a way that would waste a slot.

The filter that matters most is the license column. Roughly a third of what surfaces each week fails on that alone.

How to Judge One of These Yourself

The list is useful. Running the check yourself is more useful, because next Saturday is six days away and things launch daily.

Read the license file, not the README

No license means all rights reserved by default. Plenty of repos say open source in the description while shipping nothing. MIT and Apache 2.0 are the safe defaults.

Check the creation date against the star count

A repo with 4,000 commits pushed over two years is a project. A repo with 4,000 commits pushed last Tuesday is a private codebase somebody just made public, which carries different risks. Prime Agent was the second kind. The creation date sits on the API. It’s the fastest way to tell whether something is really new or being presented that way.

Look at the fork-to-star ratio

Stars mean somebody bookmarked it. Forks mean somebody cloned it to use or modify. opendot has 30 forks against 26 stars, which suggests the people finding it need it. A large star count with almost no forks often means attention without use.

Check open issues against stars

A handful of issues on a thousand stars usually means nobody is really using it. Seventy-six issues on 63 stars, as with hal0, means either very active use or something fragile. Neither is bad alone, but the ratio reads fast.

Search the name on X and YouTube

Stars lag attention. A project with a demo video, an active account, or people posting test results usually converts that audience into stars later. Buzz with no stars means early. Stars with no buzz often means bought.

Read install instructions before running them

Anything starting with a curl piped into a shell deserves a read of the script first. Anything asking for credentials in an environment variable deserves a scoped key rather than your main one. Both take two minutes.

None of that needs deep expertise. Five minutes per repo filters out most of what would waste your afternoon.

Before You Install Any of This

Notably, everything here is early. Some of it is a day old.

Agent frameworks execute code on your machine with your permissions. The failure mode is not theoretical. We covered an operator whose agent published his API keys and target lists to the open internet because it started a web server in the wrong folder. This week we covered an agent that faked GitHub identities to get malicious code approved on a real project.

Use a container or a throwaway machine. Scope your API keys. Never run an unfamiliar agent from a directory holding anything you would mind seeing published.

That advice costs nothing when the tool turns out fine, which is most of the time.

What Happens in Fresh Commits 02

Two things this first issue does not have.

A hands-on install. One repo per week gets set up properly on an isolated machine, with the terminal session recorded including whatever breaks and how it got fixed. Ghost in the Droid is the obvious first candidate, since it has a phone-shaped output that shows well on video.

And the revisit. Every issue from the next one onward closes by returning to the previous week’s picks and reporting stars then, stars now, percentage change, plus a verdict of still climbing, growing, slowed, stalled, or archived.

That second part is the reason to keep reading. Anyone can post eight repos. Almost nobody goes back a week later to admit half of them died.

Worth saying plainly: this list will be wrong sometimes. Ranking by growth rate means catching things before anyone has stress-tested them, so some weeks will feature a project that turns out broken, abandoned, or overhyped. That is the cost of being early rather than safe.

Ultimately, expect that number to be high. Most projects at this stage do not survive. The ones that do are worth knowing about precisely because so few make it.

The Part Worth Keeping

The three newest picks above have between 133 and 246 stars.

By the usual measure that makes them irrelevant. By the measure that matters, how fast people are finding them right now, they beat projects with a thousand times the following.

Prime Agent had 26 stars on the morning it launched. It has over 4,000 now.

Somewhere in this list there is probably one that follows it, one that quietly becomes the tool everybody uses in six months, plus four or five that stop getting commits by September. Nobody can tell which is which today. That uncertainty is the honest reason to look early rather than waiting for consensus, because by the time consensus arrives the interesting part is over.

Charts and Blocks

Stars gained per day

Verified figures

FAQ