skill-pull: Auto-load agent skills before you build

The problem

Every time I started a new project with Cursor or Claude Code, the agent would produce decent code — but miss the nuances.

It didn't know Vercel's React rendering patterns. It didn't know my Supabase auth setup. It didn't know the Stripe checkout best practices that save you hours of debugging.

I'd manually hunt for relevant Cursor rules or skill files, copy them into my project, and then start building. Every time.

There had to be a better way.

---

What is skill-pull?

skill-pull is an MCP (Model Context Protocol) server that runs alongside your AI agent. Before your agent writes any code, it:

1. Searches skills.sh semantically for skills relevant to your task
2. Installs matching skills into your project at .cursor/skills/
3. Loads their content into the agent's context

All automatically. You just describe what you want to build.

"Build a SaaS app with auth and Stripe"
→ Finds: supabase-auth, stripe-payments, nextjs-best-practices
→ Installs to my-project/.cursor/skills/
→ Agent builds with expert knowledge loaded

---

How it works

The MCP tools

skill-pull exposes four tools over the MCP stdio transport:

The auto-trigger rule

The magic is in an agent rule (auto-skills.mdc for Cursor, CLAUDE.md for Claude Code) that fires before every build task:

Before starting any build task:
1. Call recommend_skills with the task description
2. Call install_skill for each result — pass projectPath
3. Call get_skill_details to load content into context
4. Then build

The agent handles all of this transparently. You never have to think about it.

Project-level skills

Skills install into <your-project>/.cursor/skills/ — not globally. This means:

• Skills travel with the repo
• Teammates get the same context automatically
• Different projects get different skills
• Any MCP-compatible agent picks them up (Cursor, Claude Code, Windsurf, Copilot)

---

Architecture

User prompt
    │
    ▼
[auto-skills rule fires]
    │
    ▼
recommend_skills(query)
  └─► skills.sh/api/search
  └─► Returns ranked skill list
    │
    ▼
install_skill(id, projectPath)
  └─► Try npx skills add
  └─► Fallback: skills.sh/api/download → write files to disk
    │
    ▼
get_skill_details(id)
  └─► skills.sh/api/download
  └─► Extract SKILL.md → return to agent context
    │
    ▼
Agent builds with loaded knowledge

The fallback in install_skill is key — it works even if the skills CLI isn't installed, by hitting the skills.sh download API directly and writing files to disk.

---

Why I built it

skills.sh is a growing ecosystem of reusable agent skills. But discovery was manual — you had to know what to search for and run npx skills add yourself.

I wanted the agent to handle skill management the same way a senior engineer handles onboarding: read the context, figure out what knowledge is needed, and load it before starting.

skill-pull makes skills.sh feel like a background service rather than a directory you browse.

---

Try it

Add to your ~/.cursor/mcp.json:

{
  "mcpServers": {
    "skill-pull": {
      "command": "npx",
      "args": ["-y", "@rohangore1999/skill-pull"]
    }
  }
}

Then run the setup to install the auto-trigger rule:

npx @rohangore1999/skill-pull setup

Restart Cursor, start a new build prompt, and watch it pull skills automatically.

---

Links:

• GitHub
• npm
• skills.sh