skyboy.in

A fast, searchable, agent-agnostic home for portable AI skills. Each one is a self-contained SKILL.md package that hands coding agents, writing agents, memory systems, and reasoning workflows reusable, expert-level behavior.

Skyboy is the curated directory where you find a skill once and pull it into whichever agent you're already using, without copy-pasting. It ships the catalog as a website, a single dependency-free CLI binary, and an MCP server, so a skill is one command away from any tool.

---

Quickstart & Installation

The CLI

One static binary, written in Go, no runtime required:

# macOS / Linux
curl -fsSL https://skyboy.in/install.sh | sh

# Windows (PowerShell)
irm https://skyboy.in/install.ps1 | iex

Or grab a prebuilt binary straight from GitHub Releases (linux, macOS, Windows; amd64 + arm64), drop it on your PATH, and:

skyboy add <name1,name2,...>   # download into ./.skyboy/skills/ + track in ~/.skyboy/state.json
skyboy update <name>           # re-fetch and report what changed
skyboy list                    # what you have locally
skyboy list --all              # the full catalog, grouped by category
skyboy zip <name1,name2,...>   # one ZIP + a generated _CONTEXT_SUMMARY.md for ChatGPT/Claude/Gemini uploads
skyboy info <name>             # print a skill's SKILL.md
skyboy doc --print             # the docs, in the terminal
skyboy mcp --transport stdio   # the MCP server (or --transport http)
skyboy version

add installs every named skill (comma-separated) into ./.skyboy/skills/ and records it in ~/.skyboy/state.json, so list, info, update, and zip all work offline against the cache. Scoped ids work too: skyboy add @vercel/nextjs-skill. zip always writes a generated _CONTEXT_SUMMARY.md at the archive root: a short, first-class brief that tells the receiving LLM what is loaded and how to use it, which is what makes the upload-the-zip workflow actually work.

Skill format

Every skill folder (skills/<category>/<name>/) carries:

• SKILL.md with frontmatter name/description and a ## Command section holding the exact skyboy add <name> invocation
• skill.json, validated against scripts/schemas/skill.schema.json

CI runs the Go validator (skyboy validate) on every PR, and skyboy build-catalog regenerates catalog.json, including the dynamic category list: categories are derived from the top-level folders under skills/, so adding a category is just adding a folder and a PR, with no code change anywhere in the site or CLI.

Direct download (no CLI, no file server)

# degit clones just the skill folder, no git history
npx degit aijadugar/skyboy/skills/<category>/<slug> .claude/skills/<slug>

# or pull the raw SKILL.md straight from GitHub
curl -O https://raw.githubusercontent.com/aijadugar/skyboy/main/skills/<category>/<slug>/SKILL.md

---

MCP Server Setup

The Skyboy MCP server exposes the catalog as callable tools. Use the hosted endpoint for a zero-install, read-only connection, or the same skyboy binary in stdio mode if you also want install_skill (which writes to your filesystem).

Hosted endpoint (mcp.skyboy.in)

{
  "mcpServers": {
    "skyboy": {
      "type": "http",
      "url": "https://mcp.skyboy.in"
    }
  }
}

Read-only: search, preview, list, bundle. No filesystem writes.

Local stdio (install-capable)

skyboy mcp --transport stdio

{
  "mcpServers": {
    "skyboy": {
      "command": "skyboy",
      "args": ["mcp", "--transport", "stdio"],
      "type": "stdio"
    }
  }
}

Available tools

Tool	Transport	Description
search_catalog(query, category?)	remote + stdio	Fuzzy search slug, name, description, or tag.
get_skill(slug)	remote + stdio	Full SKILL.md body + skill.json metadata in one call.
list_categories()	remote + stdio	The dynamic taxonomy tree and compatible agents.
prepare_context_zip(slugs[])	remote + stdio	The exact skyboy zip bundle. stdio returns a file path; http returns a signed download URL.
install_skill(slug, target_dir?)	stdio only	Write a skill to a local folder.

One server implementation, transport selected by --transport stdio|http. The hosted endpoint serves only the read-only tools. install_skill is local-only because it writes to a filesystem and therefore requires local trust. Full config snippets for Claude Desktop and Cursor are in docs/mcp.md and on the site at /docs/mcp.

---

Documentation & Agent Install Paths

Full guides and per-agent install steps live at docs.skyboy.in.

Agent	Where it reads skills
Claude Code	.claude/skills/<slug>/
Cursor	.cursor/rules/<slug>/
Windsurf	.windsurf/skills/<slug>/
Gemini CLI	.gemini/skills/<slug>/
Codex CLI	.codex/<slug>/
ChatGPT	paste the SKILL.md into a custom GPT (or skyboy zip <slug> for a bundle)
Claude Desktop	upload the SKILL.md

---

Repository Architecture

skyboy/
├── apps/
│   └── web/                     Next.js 15 site (catalog, docs, hosted MCP endpoint)
│       ├── src/server/catalog/  In-app catalog reader (types, search, resolve, manifest)
│       ├── src/server/mcp/      Hosted read-only MCP server logic
│       └── public/install.sh    curl install script (+ install.ps1 for Windows)
├── cli/                         The skyboy Go binary: CLI + stdio MCP server,
│                                Go stdlib only, zero third-party dependencies
├── skills/                      THE catalog, one folder per skill
├── scripts/                     export-catalog, validate-skill, generate-manifest,
│                                detect-duplicates
└── docs/skill-spec.md           the canonical SKILL.md format

catalog.json at the repo root is the single shareable manifest: generated from the real skills/ tree by the Go build-catalog, and consumed by the website, the Go CLI, and the stdio MCP server.

The Go binary talks to the same raw.githubusercontent URLs and the same hosted /api/search endpoint the site serves, so the CLI works outside a checkout with no local state. Building it yourself:

cd cli
go build -o skyboy .
go test ./...

---

Contributing & Validation

Submit a skill

1. Create skills/<category>/<slug>/SKILL.md with a frontmatter name and description and a ## Command section, plus skill.json for category, tags, command, author, license, and version. Both files are validated against the schemas in scripts/schemas/. See docs/skill-spec.md for the full format.
2. Run the validation and catalog build locally (requires Go 1.24+):

   skyboy validate        # or: go run ./cli validate
   skyboy build-catalog   # or: go run ./cli build-catalog

3. Open a pull request using the skill-submission template. CI runs the Go validator against the JSON Schemas, rebuilds the catalog and fails on drift, then runs the web build and go test/go vet.

Releasing the CLI

Releases are tag-driven: push a v* tag and the release workflow cross-compiles the binary for every supported platform and attaches the artifacts to a GitHub Release, which is what the install scripts download.

Conventions

• Zero em-dashes in rendered text, docs, CLI output, and docstrings. Use commas, parentheses, or colons instead.
• Hairline borders (#c9c9c6) and pen-blue accents (#2724d1) on strokes only, never surfaces; these are the site's design tokens.

---

License

MIT. Skills carry their own licenses as declared in their metadata.json.

Built and maintained by the Skyboy project.