GITHUB HOT REPOS · SEPTEMBER 18, 2026

Hot GitHub repos
of the week.

Report by The Next New Thing.

Scroll for the report ↓

Alibaba · AI code review CLI · Apache-2.0

Ask a general AI agent to review a big pull request and it quietly skips files, so Alibaba open-sourced the reviewer it ran in-house for two years — built to be picky rather than chatty.

ELI5

A command-line tool that reads the code changes you just made and writes review comments on the exact lines that look wrong. Alibaba ran it internally for two years before opening it up.

View on GitHub →
Stars
34.6k
Language
Go
Built-in rule packs
52
GitHub repo card for alibaba/open-code-review — click to open the repo
Highlights
  • Hybrid architecture, in plain words

    Ordinary code decides which files get reviewed and which rules apply; the AI model is used only for the judgment calls inside each file.

  • Fewer false alarms, on purpose

    Its own benchmark claims higher precision than Claude Code on the same model using about one-ninth the tokens, while admitting it catches less.

  • Rule packs for 52 file types

    The repo ships per-language rules — null-pointer and thread-safety for Java, XSS for JS/TS, SQL injection for database mapper XML.

Star growth over time for alibaba/open-code-review
A clear third-party summary of what makes Open Code Review different: deterministic code picks the files and places the comments, the model only does the judging.
ayghri · AI output-style skill · MIT

Your coding assistant knows the answer but buries it under “Great question!” and three paragraphs of throat-clearing, so somebody wrote ten rules that force it to lead with the command and skip the pep talk.

ELI5

A short instruction file you install into Claude Code, Codex, Cursor and a dozen other AI assistants. It changes how they answer you: the next action first, numbered steps, no filler.

View on GitHub →
Stars
47.4k
Language
Python
Rules
10
GitHub repo card for ayghri/i-have-adhd — click to open the repo
Highlights
  • The whole thing is ten rules

    Lead with the next action, number multi-step tasks, cap lists at five items, give real time estimates, and no “Hope this helps!” closer.

  • Installs into 15 different agents

    INSTALL.md has copy-paste steps for Claude Code, Codex, Cursor, Gemini CLI, Copilot, Zed, OpenCode, Kimi, Qwen, Grok and more.

  • It graded itself and published the losses

    A blind A/B test scored it 4.47 against 4.05 for the plain prompt, then the repo names its own failed release gate and two cases it made worse.

Star growth over time for ayghri/i-have-adhd
Lays out all the rules the skill enforces and the exact install command, and ends on the line that explains the appeal: it is not really an ADHD thing, it is how most people want answers.
A well-known AI-tools channel installs the skill and shows the before-and-after on real coding answers.
A 29-second Short that explains the whole idea faster than anyone else - good for dropping in cold.
Bilawal Sidhu · live 3D intelligence globe · MIT

The satellite cockpit every spy movie fakes turns out to be buildable from public feeds, and this one runs in your browser with real planes and ships moving — no API key needed to start.

ELI5

A photorealistic 3D Earth you run on your own machine that shows real aircraft, ships, satellites, earthquakes, fires and public traffic cameras moving in near real time. You can also talk to it out loud and ask what you are looking at.

View on GitHub →
Stars
37.1k
Language
JavaScript
Live layers
15
GitHub repo card for bilawalsidhu/gods-eye-view — click to open the repo
Highlights
  • Real public feeds on one globe

    Flights from OpenSky and adsb.lol, ships via AISStream, 838 satellites from CelesTrak, USGS quakes, NASA fire detections, ~3,600 city cameras.

  • Ride inside a tracked flight

    Click any plane and the camera locks on, then drops into a cockpit view that holds the terrain under you all the way down to the runway.

  • Starts with no API keys at all

    Thirteen of the fifteen layers are keyless; optional keys add live ships, fires, Google 3D cities and an OpenAI voice mode with 28 tools.

Star growth over time for bilawalsidhu/gods-eye-view
Riley Brown's take on the whole arc - he vibe-coded Palantir, open-sourced it, and is now turning it into a startup.
This is the author's own channel, not third-party - but at 2.2 million views it IS the video everyone saw, and it is the full walkthrough from install to cockpit mode to voice control.
A third-party install route that deliberately avoids Pinokio so it does not disturb an existing AI setup.
affaan-m · agent harness toolkit · MIT

Your AI coding assistant forgets your standards the moment the chat scrolls, so this install bolts a whole engineering process onto it — plan, test, review, remember.

ELI5

A large pack of instructions, checklists and helper agents you install into an AI coding tool like Claude Code or Codex. It makes the assistant plan before it builds, write tests, review its own work from a clean slate, and keep notes between sessions.

View on GitHub →
Stars
261k
Language
JavaScript
Skills
292
GitHub repo card for affaan-m/ECC — click to open the repo
Highlights
  • Sixty-eight agents on call

    The README counts 68 specialized subagents, 292 on-demand skills and 94 command shims covering planning, review, build repair, security and docs.

  • A workflow, not a prompt

    Plan, capture a failing test, implement, review from a fresh context, verify, then save what it learned — each step leaves evidence behind.

  • Memory that moves between tools

    A local Markdown memory vault carries context and handoffs between Claude Code, Codex, Kimi and other harnesses instead of re-pasting it.

Star growth over time for affaan-m/ECC
The post that carried ECC across X - a full breakdown of the agents and skills by job, ending on the warning that installing all 286 skills at once is the fastest way to make things worse.
1.1 million views and 37,000 likes on a 106-second explainer - this is how most people heard of ECC.
An actual screen tour of the repo - what the agents folder and the skills folder look like - from back when it was at 102,000 stars.
mksglu · AI context optimizer · Elastic License 2.0

Your AI coding assistant runs out of memory and forgets what it was doing halfway through the job, so this one keeps the raw junk out of its head and stretches a session from minutes into hours.

ELI5

An add-on for AI coding tools that runs big, messy tool output inside a sandbox first, so only the short answer lands in the model’s limited working memory. It also keeps notes in a local database so the assistant remembers what it was doing.

View on GitHub →
Stars
23.4k
Language
TypeScript
Platforms
17
GitHub repo card for mksglu/context-mode — click to open the repo
Highlights
  • Sandbox the output, not the answer

    Raw data runs in an isolated subprocess and only what the script prints enters the chat. The README benchmark shows 315 KB becoming 5.4 KB.

  • Survives a context compact

    Every file edit, git action, task and error is logged to a local SQLite database, so the agent resumes instead of asking you to repeat yourself.

  • Plugs into 17 coding clients

    Claude Code, Cursor, Codex CLI, Gemini CLI, Copilot, Zed, Kiro and more, through MCP plus hooks that can block an unrouted command.

Star growth over time for mksglu/context-mode
This is the author's own release post, not a third-party reaction - but it carries the numbers from a real five-hour session: 29.6 MB of context down to 4.0 MB.
The strongest video of the whole batch - a real channel installs it, runs a live log-analysis demo, and totals up what it saved.
A 27-second version of the same pitch for anyone who does not want the six-minute one.
Tencent · RAG knowledge platform · MIT

Every company has a folder of documents nobody can find anything in, and Tencent open-sourced the system that turns that pile into something you can just ask, complete with a wiki it writes and maintains on its own.

ELI5

You upload your documents and it builds a searchable knowledge base you can talk to, answering from your own files rather than from whatever the model happened to memorize. Tencent runs it behind the WeChat Dialog Open Platform.

View on GitHub →
Stars
26.2k
Language
Go
Doc formats
13 listed
GitHub repo card for Tencent/WeKnora — click to open the repo
Highlights
  • Ask your own documents

    RAG means it finds the relevant passages in your files first, then has the model answer from those, with inline citations back to the source.

  • Writes and keeps its own wiki

    Wiki Mode has agents distill raw uploads into linked markdown pages with a knowledge graph, revision history and one-click rollback.

  • Runs entirely on your machines

    Docker Compose start, then swap in OpenAI, Gemini, DeepSeek, Anthropic or a local Ollama model, and pick your own vector database and storage.

Star growth over time for Tencent/WeKnora
WeChat's official account introduces WeKnora and links straight to the repo — this is Tencent's own launch post, not an outside reaction.
OpenAI · Codex plugin marketplace · No license

Before an AI coding agent can touch your Figma file or your Stripe account, someone has to write down how — OpenAI wrote 62 of those instruction packs and left the folder open for anyone to read.

ELI5

OpenAI’s official collection of add-ons for Codex, its coding agent. Each add-on teaches Codex how to work with one outside service — Figma, Slack, Stripe, Shopify — so you can ask in plain English instead of wiring it up yourself.

View on GitHub →
Stars
6.9k
Language
JavaScript
Plugins
62
GitHub repo card for openai/plugins — click to open the repo
Highlights
  • 62 plugins, one folder each

    The plugins/ directory holds 62, covering Figma, Notion, Slack, Stripe, Shopify, GitHub, Supabase, Vercel, Zoom, Datadog and more.

  • Skills are the moving part

    536 SKILL.md files hold the actual step-by-step instructions; Figma’s seven cover design-to-code, Code Connect and design-system rules.

  • Two marketplace lists

    marketplace.json offers 65 plugins to signed-in users; api_marketplace.json offers a narrower 49 to API-key users.

Star growth over time for openai/plugins
An independent developer opens the Codex plugin marketplace on launch day and shows what a plugin actually bundles — skills, app connections, and MCP servers.
OpenAI's own short showing plugins being shared across a workspace — official channel, so it's a product demo rather than an outside opinion.
max-sixty · Git worktree CLI · MIT or Apache-2.0

Running five AI coding agents at once means five copies of your project, and hand-juggling those copies is where it falls apart — this Rust tool shrinks the juggling to three commands, then cleans up after itself.

ELI5

A command-line tool for git “worktrees” — separate working copies of the same project that sit side by side on your disk. Worktrunk makes creating, listing, merging and deleting them one short command each, so several AI agents can each have their own copy.

View on GitHub →
Stars
7.9k
Language
Rust
Forks
272
GitHub repo card for max-sixty/worktrunk — click to open the repo
Highlights
  • One command, not three lines

    wt switch, wt list and wt remove replace the raw git incantations that make you type the branch name three times just to start.

  • Start an agent with the worktree

    wt switch -x claude -c feature-a -- 'Add user authentication' makes the branch, makes the folder, and launches Claude on that task in one line.

  • Worktrees share the build cache

    On APFS, btrfs and XFS, ten worktrees each get target/ and node_modules/ without rebuilding or copying them.

Star growth over time for max-sixty/worktrunk
Worktrunk's own author announcing it — a git worktree manager built for running AI agents in parallel. This is the author's launch post, not an outside reaction.
A 13-minute third-party review calling Worktrunk the best feature Git never got, with the whole video spent driving it on real branches.
Addy Osmani · Workflow pack for AI coding agents · MIT

Turned loose on a task, an AI coding agent takes the shortest path and skips the spec, the tests and the security pass—so a former Google engineering director wrote 25 workflows that make it slow down and show its evidence.

ELI5

A free pack of 25 step-by-step workflows you drop into an AI coding tool like Claude Code, Cursor or Codex. Each one tells the agent how a senior engineer would do that job—write the spec first, test it, prove it works—so it stops declaring victory early.

View on GitHub →
Stars
95.8k
Language
JavaScript
Skills
25
GitHub repo card for addyosmani/agent-skills — click to open the repo
Highlights
  • Nine commands, one per stage

    The pack front-ends the skills with slash commands: /spec, /plan, /build, /test, /constraints, /review, /webperf, /code-simplify and /ship.

  • A table of the excuses agents make

    Every skill ships an anti-rationalization table—the excuses agents use to skip a step, like “tests later,” each with a rebuttal.

  • One command installs all of it

    “npx skills add addyosmani/agent-skills” installs the whole set; the README also documents native setups for nine tools, Claude Code through Codex.

Star growth over time for addyosmani/agent-skills
The post that put Agent Skills in front of 420,000 people — a third-party developer explaining that AI agents left alone take shortcuts, and this pack stops them. Note the post says “from Google”: Osmani is a former Google director, and his GitHub profile still lists the old employer.
Half an hour of someone actually adding a feature to a real app with the skills, going step by step from /spec through /ship — including the part where it errors and he fixes it.
Addy Osmani himself building a habit tracker live at Microsoft Build while explaining why each skill exists — the author's own walkthrough, on someone else's stage.
Anthropic · Agentic coding tool · Proprietary

The week’s most-starred repo is “all rights reserved” and holds almost none of the product’s code—what it actually holds is a near-daily changelog and 12,452 open issues that Claude Code sorts by itself.

ELI5

This is the home base for Claude Code, Anthropic’s AI coding assistant that runs in a terminal window. The repo isn’t the program—it’s where you get the install command, read what changed, and file bugs.

View on GitHub →
Stars
146k
Language
TypeScript
Open issues
12,452
GitHub repo card for anthropics/claude-code — click to open the repo
Highlights
  • The backlog is traffic, not rot

    Roughly 11,700 open issues sit against 80,700 already closed; about 7,700 were filed in the last month and only ~350 have gone untouched since June.

  • Claude triages Claude’s bug reports

    Ten of the repo’s thirteen GitHub Actions exist to manage issues, and two of them run Claude Code itself to label and de-duplicate every new report.

  • What landed this week

    Version 2.1.275 syncs the skills and plugins on your claude.ai account into terminal sessions, and adds a ctrl+enter key that sends queued messages at once.

Star growth over time for anthropics/claude-code
Claude Code's own creator announcing Claude Mods three days ago, and the link he gives is not a blog post — it is a GitHub issue on this repo.
Nine features most people never touch — subagents, hooks, worktrees, headless mode, checkpoints — each demoed, posted ten days ago.
Half a million views of one developer's honest accounting of a thousand hours in the tool, including what stopped working for him.

First Stars

FROM THE AUDIENCE

Five agent terminals running, and the thing that breaks is your memory, not theirs.

WHAT IT IS

A macOS app that holds your Claude Code, Codex, OpenCode and Antigravity sessions the way a browser holds tabs. Panels track the decisions an agent asked about and you never answered, what it finished, what it is blocked on, and a digest of everything that happened while a tab was out of view.

View on GitHub →
Stars
4
Language
TypeScript
Started
Jul 2026
Sent in by Josh

Almost everything shipping in AI right now is aimed at managing the agents — orchestration, subagents, tool-calling — but Josh thinks the real bottleneck is the human running five or more terminals at once and losing track of which one just asked a question. Logic Loop hooks each agent's structured event stream into a side panel, so you see what actually ran without chasing tabs.

Your turn

Building something? Send it to hi@TheNextNewThing.ai

FROM THE AUDIENCE

Every Claude Code session on your machine, on one board that tells you which one is stuck.

WHAT IT IS

A read-only local web page built entirely from the files Claude Code already writes to ~/.claude: a departures board of live sessions that turns amber with the question when one stops to ask you something, a diff of every file a session touched, full transcripts, and cost per project and per week. Plain Node, zero npm dependencies, bound to localhost.

View on GitHub →
Stars
1
Language
JavaScript
Started
Aug 2026
Sent in by Jon

Jon watches the Next New Thing videos and sent in the dashboard he built. The problem in his own words: Claude Code runs in a terminal, and once you have two or three sessions going you lose track — which one is waiting for an answer, what it changed while you were away, what the week cost. This answers all three on one screen.

Your turn

Building something? Send it to hi@TheNextNewThing.ai

FROM THE AUDIENCE

Jira, GitHub, GitLab and Azure DevOps collapsed into one board that shows how engineering really runs.

WHAT IT IS

A source-available engineering cockpit that pulls Jira, GitHub, GitLab and Azure DevOps into one warehouse of development data, then reports DORA metrics, dashboards, rankings and roadmaps over it. Licensed FSL-1.1: free to run at any size forever, with the one restriction that you cannot resell it as a competing hosted service, and every release converts to Apache-2.0 two years after it ships. There is a live demo at demo.deckgauge.com with no install.

View on GitHub →
Stars
2
Language
TypeScript
Started
Jul 2026
Sent in by Yaniv

Yaniv is the founder of Codpal, a CTO-as-a-service company. After years of walking into new companies and needing a fast read on how engineering actually runs, they built Deckgauge — and the twist he offers for the show is that it was largely built by their own multi-agent system, Paperclip, using a 45,000-line STATE.md as shared memory. They have published the source so anyone can read it and run it themselves.

Your turn

Building something? Send it to hi@TheNextNewThing.ai

FROM THE AUDIENCE

A one-person business finally gets a chief of staff and a staff to hand the work to.

WHAT IT IS

An MIT-licensed folder that turns Claude or Codex into a small company: Dean, an AI chief of staff, four departments, a starter team of AI employees, and playbooks that carry multi-step work from request to finished result. Everything lives in Markdown and JSON you own, with no server and no database, and there is a non-developer install path at TrainingSites.io.

View on GitHub →
Stars
2
Language
Python
Started
Sep 2026
Sent in by James

James runs TrainingSites.io and open-sourced the system he uses himself. Campus AI OS is free and MIT-licensed, and gives anyone a desktop chief of staff plus an AI staff to do the work — no technical skills required, with everything staying local as a second brain they own.

Your turn

Building something? Send it to hi@TheNextNewThing.ai

BEFORE YOU GO

If this was useful, do three quick things.

Subscribe

New hot repos every week — don't miss the next drop.

Like

It tells YouTube to show this to more builders like you.

Comment

Tell me which repo you'll try first — or what to cover next.

THAT'S THE TOP 10 · SEPTEMBER 18, 2026

Thanks for watching.
The Next New Thing.

New hot repos every week. Subscribe so you never miss the next drop.