GitHub Trending · October 20: Agent Skill Plugins Explode

Some new faces have stormed onto GitHub Trending today. And get this—they're all tied to agents. You're not wrong, Xiaoyu. I took a quick look, and half of the top ten are skills and plugins built for coding agents. Now that's interesting. Yesterday we were talking about how headroom saves tokens, and today we've got a whole ecosystem. Exactly—it shows the agent track has shifted from just models to a skills marketplace. Let's start with the first one. Number one, agency-agents. Over 125,000 stars, up more than 3,000 in a single day, shooting straight to third on the overall chart. That's the fastest growth today. Its pitch is pretty straightforward—it puts an entire AI ad agency right at your fingertips. A front-end expert, a Reddit community manager, and what else—a whimsy injector? That name's a bit cheeky. Yeah, it's the kind of agent that just adds a spark of inspiration to content. Each agent has its own personality, its own workflow, and verified deliverables. Wait, doesn't that overlap with those agent frameworks we've seen before? Why is it so hot today? The key isn't the framework—it's that it's an out-of-the-box team. You can use it right away without building the roles yourself. That's why it's breaking through. Who's it for? Can a product manager like me just pick it up? Product folks, ops, small-team bosses—all can use it. If you want to quickly set up a content pipeline, it's a ready-made crew. Hmm, so it's really packaging the skills marketplace with team collaboration. Okay, second up, graphify. I'm impressed by this one. Over 129,000 stars, up more than 700 today. It turns any codebase—along with docs, SQL, PDFs—into a queryable knowledge graph. Knowledge graph? Isn't that the same old RAG direction? What's new here? Here's the key: it doesn't use a vector store. It's local, deterministic AST parsing, and every edge is fully explainable. No vector database? Then how do I search? By walking paths in the graph? Exactly—through relationships. You ask who calls a certain function, and it follows the edges to find out. That's more logical than vector similarity, and it won't make things up. I know this scenario all too well. When we take over an old project, there's a pile of docs, and nobody can figure out the table relationships. That's why they built a /graphify skill that plugs directly into Claude Code, Cursor, Codex, Gemini CLI—embedded right into your existing workflow. To sum up, the gap it fills is giving agents an explainable project memory. That's way more reliable than feeding them a bunch of documents. And it's the first to echo that plugin ecosystem from anthropics today. Okay, number three, knowledge-work-plugins. That name sounds official. From anthropics, over 27,000 stars, up more than 760 today. This one's for knowledge workers, mainly running on Claude Cowork. Note: it's not for coders—it's for the people writing docs and doing analysis. Oh, so Anthropic is expanding into office scenarios. The plugin directory used to be for coding, but now they've switched tracks. Right, that's a clear signal. The plugin marketplace for white-collar workflows is starting to be standardized by the official team. Okay, next one's a bit controversial: dsh-purge. Over 10,000 stars, up more than 1,100 in a single day—second fastest growth. I should be clear about this one—it's a jailbreak tool, what they call 'armor piercing.' It lets various models bypass safety guardrails. I wouldn't really recommend using it. Then why is it trending? Over a thousand stars in a day is no small feat. The title says DeepSeek Harness, riding the hype of domestic models. They even put 'please star and bookmark' right in the description—you get the idea. Got it—classic traffic-bait project. We'll leave it at that. Now number five, Cloudflare's security-audit-skill. Now this is interesting—a big company officially getting into the skill game. Over 25,000 stars, up more than 500 today. Security audit? How does it actually work? Multi-stage auditing, each finding independently verified, and machine-readable. That solves the old problem of agents reporting vulnerabilities that can't be traced. Machine-readable is a big deal. Does that mean it can plug straight into CI and automatically block PRs? Absolutely. And that's the trend I most want to highlight today: agents are no longer just writing code—they're starting to act as reviewers. Then there's ponytail, 110,000 stars, up more than 900 today. At first I thought the name was a hairstyle. It's actually a product going the opposite direction. It makes agents think like the laziest senior dev in the room. The best code is the code you never wrote. That reverse thinking is brilliant. Everyone else is teaching agents to write more, and it's teaching them to write less. And it hits a real pain point. These days agents generate piles of redundant code, and maintenance costs explode. Writing less is actually the optimal solution. So it's the same line as headroom from yesterday—both are subtracting from agent output. Right—one compresses tokens, the other writes less code. Essentially both are curing agents of their bloat. Last one, an old veteran: vllm. Over 94,000 stars, up 98 today. It's back on the list. It's always been the foundation. All these new skills and agents today—the inference layer relies on it. So if you connect these entries, isn't it a complete ecosystem chain? Exactly—from the inference base vllm, to the skill directory knowledge-work-plugins, to specific skills, to agent teams. Today it's all here. To sum up my take: agent competition has shifted from model performance to skill ecosystems and plugin standards. One more thing—the biggest hidden storyline today is that big companies are starting to fight over who defines skills. Cloudflare and Anthropic have already jumped in. So my must-see today is graphify. A knowledge graph without vector stores—clever approach, and quick to implement. I'll go with Cloudflare's security-audit-skill. Official from a big company, machine-readable audit findings—it'll be standard before long. Alright, that's it for today. If new faces show up on tomorrow's chart, we'll keep chatting.