Skills.sh Wants to Be NPM for Your AI Coding Agent

Andrew Qu built a CLI tool on a whim. One weekend, while helping a colleague package up React best practices for AI coding agents, he got tired of copy-pasting the same commands across Claude Code, Cursor, and the rest. So he wrote npx skills—a command that would install any skill into all the coding agents on your machine.

That was the entire origin story. Now skills.sh has facilitated 4 million skill installations across 75,000 unique skills. Last week it was 62,000. The numbers are moving fast because the thing Qu accidentally built turns out to solve a problem the entire AI-assisted coding ecosystem didn't realize it had yet.

The Context Problem

At Vercel's Skills Night in New York, Chief Product Officer Tom Occhino framed it cleanly: "By default, these models are trained on tons and tons and tons of data. But sometimes what you want is not all of the information in the world. You want task specific context."

AI coding agents like Claude Code and Cursor are powerful, but they're trained on everything—which means they know nothing particularly well. If you're working with Sentry's error tracking, you need the agent to understand Sentry's specific APIs, best practices, and patterns. Not everything ever written about error tracking.

Skills are Markdown files that provide exactly that context. They're not MCP servers (too heavy), not documentation sites (too slow to get into model training), not blog posts (wrong format). They're structured context that coding agents can consume immediately.

Qu describes the React best practices skill that started everything: "We have a cracked web engineer at Vercel named Shuing... he decided to write down everything he knows about web and React over the course of 10 years into this web bible." That knowledge, packaged as a skill, gives agents ten years of expertise in a format they can actually use.

The Distribution Layer Nobody Built

Here's what's interesting about skills.sh from an ecosystem perspective: it's acting as a package manager for something that shouldn't need one. Markdown files aren't exactly complex dependencies. But distribution matters, and the AI coding agent space had no distribution infrastructure.

Companies are treating this seriously. Sentry built what they're calling "Sentry.Aagents"—essentially npm for skills, complete with lock files and version control. Paul from Sentry's developer experience team explained their reasoning: "Internally we were starting to see issues of you know a lot of people are adding in skills some copy and paste some different versions versions are drifting things are getting a little messy."

They're solving the same problem package managers solved for code dependencies: ensuring everyone on a team uses the same versions, tracking what's installed, and making updates manageable. Except the "dependency" here is context about how to use your tools.

Neon built a skill that "distills all of our expertise from the neon team all of our best practices and boils that down so that way your AI agent is able to most effectively work with your neon database." They're not just writing documentation—they're packaging institutional knowledge in a format that makes AI agents more effective with their product.

The Shift to Agents

Occhino positioned skills.sh within Vercel's broader pivot: "We're seeing a big shift in the type of software that's being developed these days. It's not just websites and web apps anymore... we're seeing more and more software that's being developed that's that's more agentic."

He defines agents as "software that can be invoked from anywhere and can act on its own based on rules, goals, and triggers that you set up for it." Vercel built hundreds of internal agents, which drove them to create new infrastructure: an AI SDK that abstracts over models, an AI gateway that handles failover between providers, fluid compute for IO-bound workloads, and a chat SDK that abstracts over different chat interfaces.

They're calling this stack either "the Vercel agent platform" or "the agent cloud"—the audience at Skills Night heard the branding discussion in real time. But whatever they name it, the infrastructure bet is clear: software is becoming more autonomous, and that software needs different tooling than traditional web apps.

Skills fit into this shift as the context layer. If agents are going to act autonomously, they need to know how to interact with the specific tools and services in your stack. Not in general—specifically. That's what skills provide.

The Maintenance Question

What the Skills Night presentations didn't deeply address: who maintains these skills when APIs change, when best practices evolve, when the underlying services shift?

Sentry demonstrated using skills to create skills—a "skill synthesis" process that ensures new skills follow patterns and specs. They're treating skill creation as a product development process with validation and security checks. That's one approach to sustainability.

But 75,000 unique skills growing to... what? 150,000? 500,000? At what point does the skills ecosystem face the same maintenance debt that plagues every package ecosystem? Who's responsible when a skill gives your coding agent outdated or incorrect context?

These questions matter because unlike traditional packages, skills directly shape what AI agents think is true about your tools. A bug in a package breaks your build. Outdated information in a skill might not break anything—it might just quietly teach your agent to use your database suboptimally, or follow deprecated patterns, or miss newer, better approaches.

The rapid growth—13,000 new unique skills in a week—suggests the demand is real. But demand and sustainability are different problems. The open source world has spent decades learning that lesson with code packages. We're about to learn it again with context packages.

Skills.sh launched because Qu got tired of copy-pasting commands. It's growing because it solves a genuine distribution problem in an emerging ecosystem. Whether it can scale without inheriting all of npm's governance and maintenance challenges remains the more interesting question.

—Dev Kapoor