skill-doc-generator

MCP server that turns any documentation URL into a Claude Code skill file — and bundles the full skill-creator workflow so any agent can create, test, and iterate on skills without installing anything extra. No API key required.

Tools

The server exposes five tools. The AI agent (Claude Code, Cursor, Windsurf, or any MCP-compatible IDE) does all the reasoning:

Tool	Description
`fetch_doc_page(url)`	Fetches a documentation page, strips navigation/scripts/styles, and returns clean text plus all same-domain links.
`fetch_github_repo(repo, paths?)`	Fetches key files from a GitHub repository (README, docs, examples) via the GitHub API. More reliable than HTML scraping for repos with good inline docs. Set `GITHUB_TOKEN` env var to avoid rate limits.
`save_skill(content, skill_name, output_dir?)`	Writes the generated skill markdown to disk under `<output_dir>/skills/<skill_name>/SKILL.md`.
`read_skill(skill_name, output_dir?)`	Reads an existing skill from disk. Use before updating a skill to retrieve its current content, then call `save_skill` with the merged result.
`get_skill_creator(section?)`	Returns the bundled skill-creator instructions. No arguments → main SKILL.md. Pass a `section` URI to load a subagent file (grader, comparator, analyzer) or the JSON schema reference.

The agent decides which pages to crawl, synthesizes the content into a skill, and saves it. No external AI API calls happen inside the server itself.

MCP Resources

The server also exposes the skill-creator files as MCP resources for clients that support native resource reading:

URI	Contents
`skill-creator://SKILL.md`	Full skill-creator workflow instructions
`skill-creator://agents/grader.md`	Grader subagent — evaluates assertions against transcripts
`skill-creator://agents/comparator.md`	Blind comparator subagent — judges two outputs without knowing their source
`skill-creator://agents/analyzer.md`	Post-hoc analyzer — explains why the winning version outperformed the other
`skill-creator://references/schemas.md`	JSON schema reference for evals.json, grading.json, benchmark.json

Installation

Requires Node.js 18+. No cloning needed — install directly from npm.

Claude Code (global — works across all projects):

claude mcp add --transport stdio --scope user skill-doc-generator npx skill-doc-generator

Other agentic IDEs — add to your MCP config:

{
  "mcpServers": {
    "skill-doc-generator": {
      "command": "npx",
      "args": ["skill-doc-generator"]
    }
  }
}

Cursor: .cursor/mcp.json · Windsurf: ~/.codeium/windsurf/mcp_config.json · VS Code (Copilot): .vscode/mcp.json

If you prefer a permanent global install:

npm install -g skill-doc-generator
claude mcp add --transport stdio --scope user skill-doc-generator skill-doc-generator

Usage

Generate a skill from documentation

Once the server is registered, just ask your agent:

Generate a skill from https://laravel.com/docs/13.x/eloquent

Create a Claude Code skill from the Stripe API docs: https://stripe.com/docs/api

Build a skill from https://docs.python.org/3/library/asyncio.html and save it to ./skills

Generate a skill from https://github.com/expressjs/express

The agent will:

Fetch the initial page (or GitHub repo) and inspect available content
Decide which additional pages are worth crawling (quickstart, API reference, key concepts)
Generate the skill .md with frontmatter, key concepts, common operations, patterns, and quick reference
Save it to ~/.claude/skills/ by default (or ask you where to save it)

Update an existing skill

Update my laravel-eloquent skill with the latest docs

The agent will read the existing skill, fetch updated documentation, and save a merged version.

Create or improve a skill with skill-creator

The server bundles the full skill-creator workflow. Call get_skill_creator to get the instructions, then follow the loop: draft → test → review → improve → repeat.

Use the skill-creator to help me build a skill for X

I want to improve an existing skill — use skill-creator to run evals and iterate

The bundled skill-creator covers:

Capturing intent and writing the SKILL.md draft
Writing test cases and running them (with and without the skill) as parallel subagents
Grading outputs against quantitative assertions
Launching an eval viewer for human review
Iterating based on feedback
Optimizing the skill's description for better triggering accuracy

Output format

Generated skill files follow the Claude Code skill format:

---
name: laravel-eloquent
description: Use when working with Laravel Eloquent ORM — models, relationships, queries, scopes
---

# Laravel Eloquent

...actionable instructions for the agent...

Saved files are immediately usable as Claude Code skills.

Requirements

Node.js 18+
An MCP-compatible agentic IDE (Claude Code, Cursor, Windsurf, etc.)
No API keys needed

Project structure

src/
  index.ts                    — stdio MCP server entry point
  tools/
    generate-skill.ts         — fetch_doc_page, fetch_github_repo, save_skill, read_skill tools
  resources/
    skill-creator.ts          — get_skill_creator tool + MCP resources
bundled-skills/
  skill-creator/
    SKILL.md                  — main skill-creator instructions
    agents/                   — grader, comparator, analyzer subagent files
    references/               — JSON schema reference