beckon
Beckon blog

gstack, Modern Web Guidance, OpenSandbox: 5 Coding Agent Repos

ยท Beau Johnson

In shortI count down five GitHub repos trending for people who run coding agents: NVIDIA SkillEvaluator, OpenAI Codex Security, OpenSandbox, Chrome's Modern Web Guidance and Garry Tan's gstack. The one I'd install first is Modern Web Guidance: one command, no account, tiny, and they show their numbers.

In this video I count down five GitHub repos that were trending this week for people who run coding agents: NVIDIA's SkillEvaluator, OpenAI's Codex Security, OpenSandbox, the Google Chrome team's Modern Web Guidance and Garry Tan's gstack. I went through every one myself and ran what I could without a login or Docker. If you run Claude Code, Codex or Cursor and want to know which of these is worth your time, this is the write-up.

What I cover in this video

  • 0:46 How I picked and ranked the five repos
  • 1:31 No. 5, NVIDIA SkillEvaluator and its three tiers
  • 2:17 What NVIDIA found on its own skills, and my keyless run
  • 3:04 No. 4, OpenAI Codex Security, diff scans in CI and mock mode
  • 4:38 No. 3, OpenSandbox and the Credential Vault
  • 6:08 No. 2, Modern Web Guidance from the Chrome team
  • 7:41 No. 1, gstack and how its sprint works
  • 9:59 Which ones I'd install today, and in what order

No. 5 and No. 4: SkillEvaluator and Codex Security

First, how the board works. I pulled GitHub trending on Thursday night (Nov 12, about 7 PM PT), both the weekly and daily charts, and kept only repos that change how you actually run coding agents. Three are on the weekly chart, so they rank at the top by stars this week. Two only showed up on the daily chart, so they sit at the bottom, even though one of them is from OpenAI. Star counts move fast, so treat them as a snapshot.

SkillEvaluator (NVIDIA), 108 stars that day

SkillEvaluator answers a question I don't think enough people ask: does this skill actually help my agent? Right now we grab skills off GitHub, drop them in the skills folder and hope. It is open source under Apache 2.0, and it is part of how NVIDIA checks its own verified skills. It runs a skill through three tiers:

  1. Is it safe and built right? It checks the structure and scans the skill, pulling in SkillSpector, NVIDIA's skill scanner.
  2. Does it repeat itself? It looks for guidance that repeats or overlaps with skills you already have.
  3. Does it help? It runs your agent on test tasks with the skill and without it, inside a sandbox, and compares the two.

NVIDIA ran this on over 300 of its own skills on Claude Code and Codex. On their correctness score, agents went from 46 out of 100 without a skill to 87 with one. Those are NVIDIA's numbers on NVIDIA's skills. The token numbers are the part that stuck with me: one skill cut tokens on its task by 77%, and another more than doubled them. A skill can make your agent better and burn far more tokens at the same time, and you would never know unless you measure it.

I ran the quick check on NVIDIA's sample skill with no API key and no Docker. It scored 84.5 out of 100, a B, against a pass bar of 70. Tier one also ran without a key and passed 10 of 10 checks. Tier three needs a model and a sandbox, so that is where the real setup is.

Codex Security (OpenAI), 643 stars that day

Codex Security is a command line tool and an SDK that scans your code for security holes, and it shipped a new version this week. It checks each finding against your real source code, then you can have it validate the finding, write a patch and check whether the patch actually fixes it. It can also draft a security policy for your repo, and it builds a threat model first to do that.

For anyone running agents, the part that matters is the diff scan against main, which only looks at what changed. Your agent opens a pull request, the scan runs on just that change in CI, and you get a report before anything merges. The catch: real scans need a ChatGPT sign-in, or an API key in CI. I didn't put any keys in for this video, so I installed it locally and ran its mock mode. That produces 12 fake findings, all labeled mock, so you can test the reports and your CI before a real scan ever runs.

No. 3: OpenSandbox, an isolated box with fake keys

OpenSandbox picked up 794 stars this week. It started at Alibaba and now sits in the Agentic AI Foundation. It does one job: it gives your agent its own isolated box to run code in, on your own machines. You start it with Docker on a laptop, and the same API runs on Kubernetes when you scale up. You create a sandbox, run commands in it, write files, read them back and throw it away. There are examples for running Claude Code and Codex inside one.

This matters because your agent runs code you never read: tests, install scripts, whatever it pulled off GitHub. The feature I'd look at first is the Credential Vault. Code inside the sandbox only ever sees a fake key. When a request leaves the box, a sidecar swaps in the real one, and only for the exact sites and endpoints you set up. If something in there tries to steal your key, there is nothing real to steal.

They also say a sandbox starts in about 64 milliseconds. That is their own test, in their fast Firecracker mode on a warmed-up Alibaba Cloud cluster, one at a time, not Docker on a laptop. When they fired off 20 at once, about half got turned away. I couldn't run this one because it needs Docker and I don't have Docker on this Mac.

No. 2 and No. 1: Modern Web Guidance and gstack

Modern Web Guidance (Google Chrome), 1,341 stars this week

Modern Web Guidance is stewarded by the Chrome, Microsoft Edge and Mozilla Firefox teams, plus MDN. The problem it targets: your agent learned to code from the internet, and most of the internet is old code. Ask for a popup and it writes a pile of JavaScript when browsers have had a native dialog for years.

It is a skill of about 200 tokens. When your agent hits a web task, it runs a search command on your machine, no API key, finds the best match out of about 200 guides covering 250 modern web features, and pulls just that one guide into context: the pattern, the gotchas and a fallback for older browsers.

Google tests Claude Code, Codex and Antigravity on tasks in its own test web apps, with and without the guidance. Claude Code on Opus 5.5 went from 57% to 88%, and Codex went from 54% to 82%. Those are Google's own numbers, with full breakdowns on a public dashboard (if vendor-run scores are new to you, I wrote up how coding agent benchmarks work). On a demo page I wrote full of old habits, its legacy scan found eight spots to modernize across seven guides. It didn't catch everything: my first version, with an old jQuery-style popup, got exactly one hit. It only flags patterns it has rules for.

It also sends anonymous usage data to Google, like the searches your agent makes, not your prompts. One environment variable turns that off:

DISABLE_TELEMETRY=1

gstack (Garry Tan), over 1,700 stars this week

gstack comes from Garry Tan, President and CEO of Y Combinator, who calls it the open-source software factory he uses every day. It is at over 141,000 stars total for a repo that started in March. It is a set of Claude Code skills, all slash commands, all markdown, and each one plays a role on a team. The description says 23 tools, but I counted 62 skill folders in the repo.

  • Office hours: the YC partner who pushes back on your idea before you write code.
  • Plan CEO review: rethinks the scope.
  • Designer: hunts for AI slop.
  • Review: the staff engineer looking for bugs that pass CI and break in production.
  • QA: opens a real browser and clicks through your app.
  • CSO: the security officer.
  • Ship: runs the tests and opens the pull request.

The point is that these run as one sprint: think, plan, build, review, test, ship, reflect. Every step reads what the step before it wrote down. Garry says he runs 10 to 15 of these sprints at once, and in his words:

Without a process, 10 agents is 10 sources of chaos.

It has guardrails too. Careful warns you before an rm -rf or a force push, and freeze locks edits to one folder. The repo says it installs in about 30 seconds, but it goes into your Claude Code skills folder, it needs Bun, and Claude Code is the only agent it fully certifies. Codex and Cursor are listed as experimental.

How the five fit together, and what I'd install first

If you run a few agents at once (the kind of setup Beckon is a workspace for), think of these as one line. gstack gives each agent a role and a process. Modern Web Guidance keeps the agent writing your frontend from coding like it's 2015. OpenSandbox gives the agents that run code an isolated box with fake keys. Codex Security can check every diff before it merges. SkillEvaluator checks the skills you hand all of them, including the first two on this list.

My install order:

  1. Modern Web Guidance first. One command, no account, tiny, and they show their numbers.
  2. Then gstack, starting the way Garry says: office hours, review and QA, and stop there.
  3. Codex Security in CI on your agents' pull requests.
  4. OpenSandbox and SkillEvaluator for when you start letting agents run on their own.

For more picks in this vein, see my earlier roundup of trending repos that make coding agents safer to run.

Key takeaways

  • The repo I'd install first isn't number one. Modern Web Guidance wins on effort: one command, no account, about 200 tokens.
  • Measure your skills. In NVIDIA's own results, one skill cut tokens by 77% and another more than doubled them.
  • gstack's value is the process: each role reads what the previous step wrote, which matters once you run more than one agent.
  • OpenSandbox's Credential Vault means code in the box never holds a real key, only the calls you allowed go through.
  • Every benchmark here is the vendor's own. I only ran SkillEvaluator's keyless checks, Codex Security's mock mode and Modern Web Guidance's legacy scan myself.

FAQ

What is gstack?

gstack is Garry Tan's open-source set of Claude Code skills, all slash commands in markdown, where each skill plays a role such as office hours, CEO review, designer, staff engineer, QA, security officer and ship. They run as one sprint where every step reads what the previous step wrote. It needs Bun, and Claude Code is the only agent it fully certifies.

Does Modern Web Guidance send data to Google?

Yes. It sends anonymous usage data, like the searches your agent makes, but not your prompts. Setting DISABLE_TELEMETRY=1 turns it off.

What does NVIDIA SkillEvaluator check?

It runs a skill through three tiers: whether it is safe and well formed, whether its guidance repeats or overlaps with skills you already have, and whether it helps, by running the same tasks with and without the skill in a sandbox. The first checks run without an API key, but the third tier needs a model and a sandbox.

Can I try Codex Security without signing in?

Only in mock mode. Real scans need a ChatGPT sign-in or an API key in CI, while mock mode produces 12 fake findings, all labeled mock, so you can test the reports and your CI setup first.

Run your agents in Beckon

A workspace for your coding agents. Get more work done in less time.

Download BeckonSubscribe on YouTube
Full video transcript

0:00Gary Tan, the CEO of Y Combinator, open sourced what he calls his exact claude code setup, and it still picked up over 1,700 stars this week. Google's Chrome team says their skill takes Claude code from 57 to 88% on their own tests. And Nvidia built a tool that tells you if a skill actually helps your agent or just burns more tokens. But the one I'd install first isn't number one. So, I'm counting down five repos trending right now for people who run coding agents from Nvidia's skill tester up to Gary's setup. I went through everyone myself and ran what I could without a login or Docker. Let's jump right into it. So, how did I pick these? I pulled GitHub trending on Thursday night, the weekly charts and the daily charts, and I only

0:46kept repos that change how you actually run coding agents. Three of these are on the weekly charts, so they go at the top ranked by stars this week. Two only showed up on the daily chart, so they sit at the bottom of the board, even though one of them is from OpenAI. Star counts move fast, so these are the numbers from Thursday night. That's the whole scoreboard. Number five. Number five is Skill Evaluator from Nvidia. 108 stars today. And it answers a question I don't think enough people ask. Does this skill actually help my agent? Because right now we're all grabbing skills off GitHub, dropping them in the skills folder, and just hoping. Skill evaluator is open- source Apache 2.0, and it's part of how Nvidia checks its own

1:31verified skills. So, Skill Evaluator runs a skill through three tiers. Tier one, is it safe and is it built right? It checks the structure and it scans it and it pulls in skill spectre, Nvidia's skill scanner, which was number one on our last top five. Tier two, does the guidance repeat itself or overlap with skills you already have? And tier three is the part I like. It runs your agent on test tasks with the skill and without it inside a sandbox and compares the two. And when Nvidia ran that on its own skills, one of the results went the opposite way of what you'd think. Every night, Night Crew researches what your market is talking about, writes the script, fact checks it, voices it in your voice, and animates it. In the morning, you approve it and it posts to YouTube, Instagram, Tik Tok, and

2:17Facebook. Nightcrew. Nightcrew.studio. So, Nvidia ran this on over 300 of their own skills on Cloud Code and Codeex. On their correctness score, the agents went from 46 out of 100 without a skill to 87 with one. Those are Nvidia's numbers on Nvidia's skills, so take that for what you will. Then, I got to the token numbers. One skill cut the tokens on its task by 77%. Another one more than doubled them. So, a skill can make your agent better and burn way more tokens at the same time. And you'd never know unless you measure it. So, I ran the quick check on Nvidia's sample skill. No API key, no Docker. It scored 84 12 out of 100, a B, and the bar to pass is 70. I ran tier

3:04one too, also with no key, and it pulled in skillspector and passed 10 out of 10 checks. Tier three is the one that needs a model and a sandbox. So, that's where the real setup is. Number four is codec security from OpenAI. 643 stars today, and it shipped a new version this week. It's a command line tool and an SDK that scans your code for security holes. You type codec security scan, and it goes through the repo. So, how does it actually work? It checks each finding against your real source code and then you can have it validate the finding, write a patch and check whether that patch actually fixes it. It can even draft a security policy for your repo and it builds a threat model first to do it. And for anyone running agents, the flag that matters is diff scan with-d

3:52against main and it only looks at what changed. So, your agent opens a pull request. The scan runs on just that change in CI and you get a report before anything merges. Now, the catch. Real scans need you to sign in with your chat GPT account or an API key in CI. And I didn't put any of my keys in for this video. So, I installed it locally and ran its mock mode. It spits out 12 fake findings, all labeled mock, so you can test the reports and your CI before a real scan ever runs. Number three is Open Sandbox. 794 stars this week. It started at Alibaba and now it sits in the Agentic AI Foundation. It does one job. It gives your agent its own isolated box to run code in on your own

4:38machines. Start it with Docker on your laptop and the same API runs on Kubernetes when you scale up. You create a sandbox, run commands in it, write files, read them back, and throw it away when you're done. They even have examples for running claude code and codecs inside one. So why does that matter? Your agent runs code you never read. Tests, install scripts, whatever it pulled off GitHub. Inside a sandbox, that code is walled off from your real machine. And the feature I'd look at first is the credential vault. Your agent needs to call GitHub or some API, so it needs a key. Normally, you just hand it the key. With the vault, the code inside the sandbox only ever sees a fake key. When the request leaves the box, a sidecar swaps in the real one and only for the exact sites and endpoints

5:23you set it up for. So if something in there tries to steal your key, there's nothing real to steal. It can still make the calls you allowed. It just never holds the key. That's really cool. They also say a sandbox starts in about 64 milliseconds. That's their own test with their fast firecracker mode on a warmed up Alibaba cloud cluster one at a time, not Docker on a laptop. When they fired off 20 at once, about half got turned away. So, it's fast, but it's not magic. And I couldn't run this one today because it needs Docker, and I don't have Docker on this Mac. Yeah, I know. Number two is modern web guidance from the Google Chrome team. 1341 stars this week, and it's stewarded by the Chrome, Microsoft Edge, and Mozilla Firefox

6:08teams, plus MDN, three of the big browser makers in one repo. Your agent learned to code from the internet and most of the internet is old code. So you ask for a popup and it writes a big pile of JavaScript when browsers have had a native dialogue for years. So this is a skill. When your agent hits a web task, it runs a search command and that search runs right on your machine. No API key. It finds the best guide out of about 200 guides covering 250 modern web features. And then it pulls just that one guide into the context window. The pattern, the gotchas, and a fallback for older browsers. The skill itself is about 200 tokens. So it stays tiny until it needs something. And they test it. They run

6:55Claude code, codeex, and anti-gravity on tasks in their own test web apps with the guidance and without it. Claude code on Opus 5.5 went from 57% to 88%. Codeex went from 54 to 82. Those are Google's own numbers, but they put the full breakdowns on a public dashboard. I tried it on a little demo page I wrote full of old habits, and its legacy scan found eight spots to modernize across seven guides, but it didn't catch everything. My first version with an old jQuery style popup got exactly one hit. It only flags the patterns it has rules for. It also sends anonymous usage data to Google, like the searches your agent makes, not your prompts. You can turn it off with one environment variable. And

7:41number one with 174 stars this week is GStack from Gary Tan, the president and CEO of Y Combinator. He calls it the open-source software factory he uses every day. It's at over 141,000 stars total, which is insane for a repo that started in March. So, what is it? It's a set of Claude code skills, all slash commands, all markdown, and each one plays a role on a team. The description says 23 tools, but I counted 62 skill folders in the repo today. It keeps growing. Office hours is the YC partner who pushes back on your idea before you write any code. Plan CEO Review rethinks the scope. There's a designer that hunts for AI slop. Review is the staff

8:27engineer looking for the bugs that pass CI and then break in production. QA opens a real browser and clicks through your app. CSO is the security officer. And ship runs the tests and opens the pull request. Now, here's what this is all about. All of these run as one sprint. Think, plan, build, review, test, ship, reflect. Office hours writes a design dock and the CEO reviews it. Auto plan writes a test plan and QA picks it up. review catches a bug and ship checks that it's fixed. Every step reads what the step before it wrote down. And that's why it matters once you run more than one agent. Gary says he runs 10 to 15 of these sprints at once. And in his words, without a process, 10 agents is 10 sources of chaos. It comes

9:14with guard rails, too. Careful warns you before an RM-rf or a force push and freeze locks edits to one folder. The repo says it installs in about 30 seconds, but it does go into your Claude code skills folder and it needs bun and claude code is the only agent it fully certifies. Codeex and cursor are listed as experimental. So, how do these five fit together if you run a few agents at once? Think of it as one line. GStack gives each agent a role and a process. Modern web guidance keeps the agent writing your frontend from coding like it's 2015. Open Sandbox gives the agents that run code their own isolated box with fake keys. Codeex security can check every diff before it merges. And

9:59skill evaluator checks the skills you hand all of them including the first two on this list. So which one would I install today? Modern web guidance first. It's one command, no account. It's tiny and they show their numbers. Then GStack, but I'd start the way Gary says, office hours, review and Q&A, and stop there. Codeex security goes in CI on your agents pull requests and open sandbox and skill evaluator are for when you start letting agents run on their own which is where all of this is heading. Okay, so which of these five is going into your setup first? And did I miss one that should have made the board? Let me know in the comments. And if you found a repo here you hadn't heard of, give the video a like. It helps other people running agents find it. Have a blessed day.