Cloudflare Security Audit Skill, Graphify and 3 More Trending Repos
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I count down five more free GitHub repos that blew up last week, ranked by the stars they gained on GitHub Trending. For each one I explain what it is, how it works step by step, why it took off and what to watch out for. Then I tell you which ones I'd actually install. It's for anyone using Claude Code or Codex who wants to know which trending tools are worth a look.
Quick ground rules: Orca and HyperFrames are still on the chart, but I covered them last week, so I skipped them. I also left out OpenCode because you probably already know it.
What I cover in this video
- 0:00: Part two of the trending repos series, and the ground rules
- 0:45: #5 Ruflo, formerly Claude Flow, with my update 14 months after I first covered it
- 3:06: #4 AI Memory, one memory shared by Claude Code, Codex and 30+ other agents
- 5:22: #3 ComfyUI, node-based AI image and video on your own computer
- 7:40: The official Comfy MCP for Claude Code, and ComfyUI's security warning
- 8:27: #2 Graphify, which turns your codebase into a knowledge graph
- 10:43: #1 Cloudflare's security-audit-skill, where agents try to refute each other's bugs
- 13:00: Which ones I'd actually install
#5 Ruflo and #4 AI Memory: harnesses and shared memory
Ruflo gained almost 3,000 stars last week and has around 80,000 in total. It used to be called Claude Flow. It's a free add-on that wraps Claude Code in a big harness. When you run its emit command, it writes about 100 agent files, a set of hooks and a huge CLAUDE.md into your project. It also plugs in an MCP server with over 300 tools. On every turn, hooks route your task, pull in memory and log what happened. Then there's the swarm: a queen agent with a team of workers. The thing to understand is that the workers are Claude Code's own subagents. Ruflo mostly keeps the books.
Things to be wary of: an independent audit in April called it "99% theater," and the project's own audit in May found that one of its speedup numbers came from a random-number function. To be fair, that same audit said the learning loop is real, and a lot has been patched since. In July it had a 10/10 security bug in its Docker setup, which was fixed within a day. In the project's own demo, Claude Code warns that Ruflo's CLAUDE.md is over the size limit. If you want to try it, start with the plugin in a scratch project.
AI Memory gained more than 3,000 stars last week and has over 13,000 total. Here's the problem it solves. You hit your limit in Claude Code, open Codex in the same folder, and Codex knows nothing about the last few hours of work. Its creator, Fabio Akita, built the fix in Rust, and over 1,300 of the commits are co-authored by Claude. It works in four steps:
- Capture: hooks send every prompt and tool call to a small server on your machine. Known API key formats get stripped out first.
- Consolidate: when the session ends, it writes a summary page in plain markdown inside a git folder. No AI model is needed for this step.
- Recall: the next session, in any supported agent, starts with a short briefing and can search the memory with MCP tools.
- Handoff: the old session leaves a note with open questions and next steps, and the next agent claims it exactly once.
The project lists over 30 agents, including Claude Code, Codex, Cursor and Gemini CLI. Things to be wary of: every prompt and tool output lands in a local database, and by default it imports your old agent history too. The project says its secret filter is best effort and there's no encryption at rest. Akita himself says that if you only use one agent, you probably don't need it.
#3 ComfyUI: AI image and video on your own machine
ComfyUI gained almost 5,000 stars last week and has more than 140,000 total. It's a free app for running AI image and video models locally. Instead of typing into a chat box, you wire boxes together on a canvas: load a model, write a prompt, sample, decode, save. Every box is a small Python node. When you hit run, the whole graph goes as JSON to a server on your machine. ComfyUI only runs the nodes that changed and caches the rest, which is why tweaking feels fast. It also moves model weights between your graphics card, RAM and disk so big models fit on normal cards. Every image it saves has its whole workflow baked inside, so you can drag the image back onto the canvas to get the exact graph back.
It's been around since early 2023, so the jump last week came from several things at once: day-one support for two new open image models, a new release, and $25,000 in prizes for building an open-source Higgsfield. For agent builders, the big news is that there's an official Comfy MCP, so Claude Code can build and run workflows for you.
Things to be wary of: custom nodes are Python with full access to your computer. In January, a fake upscaler node in Comfy's own registry delivered info-stealing malware on Windows. Keep ComfyUI on localhost, which is the default, and only install nodes you trust. You'll also need a real graphics card, and you should check each model's license before you sell anything you make with it.
#2 Graphify and #1 Cloudflare's security audit skill
Graphify gained almost 7,000 stars last week. You type /graphify in Claude Code and it turns your project into a knowledge graph. Every function, class, file and import becomes a node, and every call becomes a line between them. Then your agent can ask the graph something like "what connects login to the database" instead of grepping file after file. The package name has two y's at the end, and the README warns that other packages with similar names aren't theirs:
graphifyy
It parses your code locally with tree-sitter, with no AI and no internet, so that part is free. Docs and PDFs are optional, and your agent's model reads those, so that part does cost tokens. Graphify groups everything into subsystems and saves a clickable map and a report. Every line in the graph is labeled as either read from the code or inferred, so you can tell a fact from a guess. Things to be wary of: its headline "71.5x fewer tokens" number compares against reading every file, and Claude Code doesn't do that. In one tester's real Claude Code sessions, every run cost more with Graphify and gave the same answers. Install it to understand your code, not to cut your bill.
At #1 is Cloudflare's security-audit-skill, which gained just over 15,000 stars last week. It's basically a folder of markdown files that you install into Claude Code or Codex. Tell your agent "Security audit this codebase" and it works like a small security team. Read-only recon agents map your code. Hunter agents each take one slice and have to name the attacker, the input and the exact line of code. Then every possible bug goes to a fresh agent with this prompt:
You did not write this candidate. Try to refute it.
The findings are checked against a strict format, and you get a report with fixes. The skill never edits your code. Cloudflare's blog explains why it's built this way: a hunter allowed to grade its own homework will confidently validate everything, and it may even edit the code so its own exploit works.
Things to be wary of: one tester found that quick mode cost about $30 a run, versus about $2 for a single plain agent, and both caught the same share of bugs. A clean report doesn't mean you're safe, because the skill doesn't check your dependencies. One user applied the suggested fixes and more than a dozen of them removed real features, so review the fixes like you would any pull request.
Key takeaways
- Security audit skill: run quick mode on one folder of your own code, and review every fix before you merge it.
- Graphify: worth it on a big project so you can see how everything connects. Don't install it expecting lower token costs.
- AI Memory: useful if you switch between Claude Code and Codex. Keep in mind it records everything locally without encryption. More on that kind of setup in Claude Code vs Codex.
- ComfyUI: a good choice if you have a real graphics card and want to make images for free. Keep it on localhost and only install nodes you trust.
- Ruflo: start with the plugin in a scratch project, and only if you really want the big harness for your AI agents.
FAQ
What is Cloudflare's security-audit-skill?
It's a free skill, basically a folder of markdown files, that you install into Claude Code or Codex. It turns your agent into a small security team, and every possible bug goes to a fresh agent that tries to refute it. It writes a report and never edits your code.
Does Graphify reduce Claude Code token costs?
Not reliably. Its 71.5x figure compares against reading every file, which Claude Code doesn't do. In one tester's real Claude Code sessions, every run cost more with Graphify and gave the same answers. Install it to understand your code, not to cut your bill.
What happened to Claude Flow?
It's now called Ruflo. It wraps Claude Code in a harness of agent files, hooks and an MCP server with over 300 tools, but Claude Code's own subagents do the actual work. My advice is to try the plugin in a scratch project first.
Is AI Memory safe to use?
It records every prompt and tool output in a local database, and by default it imports your old agent history. The project says its secret filter is best effort and there's no encryption at rest, so go in knowing that.
Chapters
Transcript
Read the full transcript
0:00Last week's GitHub repos video got nearly 90,000 views in its first five days, so we're doing it again. In this video, we're counting down [snorts] five more free GitHub repos that blew up last week, ranked by the stars they gained. One turns your coding agent into a whole security team, and it's from Cloudflare. One turns your entire codebase into a map your agent can actually read. One lets Claude Code and Codeex share the same memory. and one of them I made a video on over a year ago and the update is not what I expected. Plus, which ones I'd actually install. Let's jump right into it. Quick ground rules. Two of last week's picks, Orca and Hyperframes, are still on the chart, so I skipped them.
0:45Same with Open Code, which you probably already know. So, number five is Ruflow. Almost 3,000 stars last week and around 80,000 total. And if that name sounds new, this might ring a bell. It used to be called Claude Flow. 14 months ago, I made a video on it. So, this is the honest update. So, what is it exactly? It's a free add-on that wraps Claude code in a giant harness. Their own line says it best. Cloudflow coordinates. Claude Code creates. Here's how it actually works. Step one, you run the emit command and it writes about a 100 agent files, a stack of hooks, and a huge claw.md into your project. Step two, it plugs in an MCP server with over 300 tools. Step three, hooks fire on
1:32every turn. They route your task, pull in memory, and log what happened so it can learn. And then step four, the swarm, a queen agent and a team of workers. But here's the thing to understand. The workers are Claude Code's own sub agents. Ruflow mostly keeps the books. So why last week? No single launch that I could find. They just kept shipping. 29 releases in about a week. Here's something to be wary about. So, one independent audit in April called it 99% theater, and the project's own audit in May found one of its speedup numbers was generated by a random number function. To be fair, that same audit said the learning loop is real, and a lot has been patched since. Back in July, it had a 10 out of 10 security bug in its Docker setup. It was
2:18fixed within a day. And in their own demo, Claude Code warns that Ruflow's claw.md is over the size limit. Come on now. And here's my favorite part. Just 50 stars behind it last week was a repo that's one claw.md file with four rules based on Andre Carpathy's notes on how coding agents mess up. It has almost three times the stars. So if you try RuFlow, start with the plugin in a Scratch project. Real quick, every video I make is built in code with Claude Code. Inside Shipping School Premium, I show you how step by step to turn your product into videos that get views, plus new courses every day, 500 plus hours of replays, live calls every week, and a Q&A with me every Saturday. Join free link below. Number four is AI memory.
3:06More than 3,000 stars last week and over 13,000 total. So, here's the problem it fixes. You're deep into a task in Claude Code. You hit your limit and you open Codeex in the same folder. In the creator's own words, "Codex knows nothing about the last four hours of discussion." So, the creator is Fabio Akita, a Brazilian developer who's been programming since the early '9s. His workaround was asking Claude to write a handoff file and then asking Codeex to read it every time. So, he built his own fix in Rust. And here's the funny part. Over 1,300 of the commits are co-authored by Claude, a memory tool for Claude, built with Claude. Here's how it works. Four steps. Capture, consolidate,
3:51recall, and handoff. Step one, capture hooks inside your agent quietly send every prompt and tool call to a small server on your own machine, and it strips out known API key formats first. Step two, consolidate. When the session ends, it writes a summary page in plain markdown inside a git folder. No AI model needed. Step three, recall. The next session in any supported agent starts with a short briefing and it can search the memory with MCP tools. And then step four, the handoff. The old session leaves a note with the open questions and next steps. And the next agent claims it exactly once. They list over 30 agents. Claude Code, Codeex, Cursor, Gemini CLI, Grock Build, you
4:36name it. So why now? On October 15th, Aikido went on Flow podcast, one of Brazil's biggest podcasts, and he talked about it. When his Claude plan runs out, he says, "Am I going to pay for extra tokens?" No. So he opens Codeex, and when Codex runs out, he moves to Kimmy. All sharing one memory. A few days later, the stars took off. It took 5 months to reach 10,000 stars. and then about a week to add over 3,000 more. So, funny enough, I built the same idea into Becken, my own agent setup. Every agent leaves a handoff note for the next one. So, I get it. I get it. Here's something to be wary about. So, it records your sessions. That's the whole point. But every prompt and tool output lands in a local database, and by default, it
5:22imports your old agent history, too. The project itself says the secret filter is best effort, not a guarantee, and there's no encryption at rest. And if you only use one agent, Akita himself says you probably don't need anything extra. Number three is Comfy UI. Almost 5,000 stars last week and more than 140,000 total. Now, this one isn't new. It's been around since early 2023. So, why did it blow up last week? We'll get there. So, what is it? It's a free app for running AI image and video models on your own computer. You don't type into a chat box. You wire little boxes together on a canvas. Load a model, write a prompt, sample it, decode it, save it. That wiring diagram is your workflow.
6:08Here's how it actually works. Step one, every box is a small Python node with inputs and outputs. Step two, when you hit run, the whole graph gets sent as JSON to a server running on your own machine. Step three, it only runs the nodes it needs and it caches the rest. So if you change one setting at the end, only that part runs again. That's why tweaking feels so fast. And then step four, the memory juggling. These models are huge. So it moves the weights between your graphics card, your RAM, and your disc, so big models fit on normal cards. And my favorite detail, every image it saves has the whole workflow baked inside. drag the picture back onto the canvas and the exact graph comes right back. So, according to Comfy, it started with one developer in
6:54Quebec City who goes by Comfy Anonymous and he built the first version in about 2 weeks. In April, Comfy raised $30 million at a $500 million valuation and he still has about 55% of all the commits. So, why last week? It's a pileup. Two new open image models landed with day one support. Flux 3 Klein and Moss Image. A new release shipped and Comfy put up $25,000 in prizes to build an open- source Higsfield because Comfy says Higsfield promised to open source its platform and didn't. For about 3 days, the stars came in five times faster than normal. And for us agent people, this is the part I'm excited about. There's an official Comfy MCP, so Claude Code can build and run workflows
7:40for you. And the wiring for a local Comfy agent just landed in the code. The agent itself is still cloud only and Comfy says local support is coming soon. Here's something to be wary about. So, custom nodes are just Python with full access to your computer and Comfy's own security policy says so. In January, a fake upscaler node in their own registry delivered info stealing malware on Windows. So, keep it on local host, which is the default, and only install nodes you trust. You'll also want a real graphics card. And every model has its own license. So check it before you sell anything you make. Number two is Graphify. Almost 7,000 stars last week and more than 135,000 total. So what is it? You type /graphify in cloud code and
8:27it turns your whole project into a knowledge graph. Every function, class, file, and import becomes a dot and every call becomes a line. Then your agent asks the graph what connects login to the database. instead of grepping through file after file. Here's how it works. Step one, install it. The package is graphify with two Y's at the end, and their readme warns that the other ones aren't theirs. Step two, it parses your code locally with Treesitter. No AI, no internet, free. Step three, docs and PDFs are optional. Your agents model reads those, so that part costs tokens. Step four, it merges everything into one graph, groups it into subsystems like off and billing, and saves a clickable map and a report. And then step five,
9:12your agent queries it, find the path between two things, explain a function, and a hook nudges your agent to check the graph before it starts grepping. And I really like this part. Every line in the graph is labeled, either read straight from the code or inferred, so you can tell a fact from a guess. So the story is wild. Safi Shamsy saw Andre Karpathy's post about LLM knowledge bases and he says he built the first version that weekend. By their count, it hit a 100,000 stars in under 4 months and now it's a Y Combinator company. And YC's page lists the team size as one. One. So why is it still adding 7,000 stars a week? No single viral post, three releases last week, a steady drum beat of creator videos, and GitHub
9:58trending feeding itself. Here's something to be wary about. So, their best number, 71 and a half times fewer tokens, compares against reading every single file. Cloud code doesn't do that. It gs. One tester ran real cloud code sessions and every run cost more with Graphify with the same answers. And here's the irony. A tool sold on saving tokens had a hook that burned about 650,000 tokens in one project just telling Claude to use it. That's fixed now. So install it to understand your code, not to cut your bill. And number one is Cloudflare's security audit skill. Number one on GitHub trending last week with just over 15,000 new stars and more than 47,000 total. And
10:43it's basically a folder of markdown files. No app, no model, 14 commits. So what is it? It's a skill you install into cloud code or codecs. You say, "Security audit this codebase and your agent turns into a small security team." Here's how it actually works. Step one, recon. Several readonly agents map your code at the same time. What the product is, who's trusted, and where outside input comes in. Step two, the hunt. Hunter agents each take one slice, and each one has to name the attacker, the input, and the exact line of code. Step three, and this is the whole point. Every possible bug goes to a fresh agent that didn't find it. And the prompt is literally, "You did not write this candidate. Try to refute it. The agent
11:29that checks a finding is never the agent that found it." Step four, the findings get checked against a strict format. And then step five, you get a report with the fixes. It never touches your code. So why build it that way? Cloudflare's own blog explains it. If a hunter is allowed to grade its own homework, it will confidently validate everything it outputs. And the agent will edit the source code so its own exploit works, then triumphantly report the bug it just created, which is insane. So this skill is the seed of Cloudflare's internal bug hunter. They say that bigger system scanned 128 repos and sent over 7,000 findings to their teams. But those numbers are the internal system, not this repo. So why last week? Honestly,
12:14nobody seems to know what lit the fuse. It was a week full of AI security news, and once it hit the top of trending, trending kept feeding it. Here's how I'd use it. We have a Chrome extension called Wind Snap running on all of our computers in the resale business. I'd run the quick mode on just that folder. Here's something to be wary about. Tokens. One tester found the quick mode cost about $30 a run versus about $2 for one plane agent, and both caught the same share of the bugs. A clean report doesn't mean you're safe because it doesn't check your dependencies. One user applied the fixes and over a dozen of them removed real features. So review the fixes like any pull request and people have filed security bugs against the security skill. A security skill with security bugs. So which ones would
13:00I actually install? The security audit skill in quick mode on one folder of your own code. Graphify if your project is big to see how it all connects. AI memory if you bounce between cloud code and codecs. Comfy UI if you have a real graphics card and you want to make images for free. And rooflow start with the plugin in a scratch project and only if you really want the big harness. So now I want to hear from you. Did I miss a repo that should have made the cut? Drop it in the comments and it might end up in part three. And if even one of these five ends up in your setup, leave a like on your way out. Have a blessed day.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Trending GitHub Repos for Claude Code: T3 Code, claude-mem, ECC
The five repos that gained the most GitHub stars last week were T3 Code, claude-mem, ECC, Superpowers and TradingAgents. I'd install Superpowers first, try T3 Code if you pay for two or more agents, take only a few pieces of ECC, hold off on claude-mem, and read TradingAgents' code without trading with it.
Top Trending GitHub Repos: VoiceStudio, ai-job-search, Orca, Ghidra
The five repos right behind REA on GitHub Trending this week are VoiceStudio (free local voice cloning), ai-job-search (a Claude Code job hunt), Orca (a desktop app for running many coding agents), Ghidra (the NSA's reverse engineering tool), and HyperFrames (HTML to MP4 video).
Claude Code GitHub Repos: Manim, Remotion, Blender MCP and More
Five free repos take Claude Code past writing code. Manim and Remotion Skills make videos, Lightpanda is a browser built for agents, MCP for Blender builds 3D scenes, and agent-desktop lets an agent use your whole computer. I'd install MCP for Blender first, and I'd only run agent-desktop in a VM or a separate user account.
