MCP 服务器

Skillselion MCP server

Enables AI agents to search a curated directory of Claude Code agent skills, MCP servers, and plugin marketplaces ranked by community signal.

README

Augment your coding agent with thousands of battle-tested skills. It searches <a href="https://skillselion.com">Skillselion</a> — a curated directory of Claude Code skills, MCP servers & marketplaces ranked by real installs — and pulls a skill's actual instructions to apply mid-task. So your agent works from proven patterns, not guesses.

🧩 Works with Claude Code · Claude Desktop · Cursor · Codex · and any MCP client

Real run against the live catalog: <code>search_skillselion</code> → <code>load_skill</code> pulls the skill's real <code>SKILL.md</code> + bundled files into a temp folder → your agent follows it like an installed skill.

✨ What it does

Skillselion indexes thousands of Claude Code skills, MCP servers and plugin marketplaces, ranked by real community signal (installs + GitHub stars). This server turns that directory into a live capability for your agent: it can search for the right skill and load_skill to dynamically load it mid-task — the real SKILL.md into context, plus any bundled scripts / references / templates materialized into a temp folder — so it works from the skill exactly like an installed one, instead of improvising "best practices." Find the skill → load it → follow it.

🔧 Tools

Tool	What it does
🔍 `search_skillselion`	Search by keyword or task (`postgres`, `code review`, `playwright`); optional `type` filter (`skill` / `mcp` / `marketplace`). Returns name, type, installs, stars, repo, an install command, and the listing URL.
📥 `load_skill`	The dynamic load. Pulls a skill's real `SKILL.md` into context and materializes its whole directory (scripts, references, templates) into a temp folder, so the agent uses it like an installed skill — read/run the bundled files included. Pass an `id` from search, or a `query`.
🏆 `top_skillselion`	The leaderboard — "what are the best Claude Code skills / MCP servers right now." Skills rank by installs; MCP servers & marketplaces by GitHub stars.

Read-only & private. The server only issues GET requests to the public Skillselion catalog API. No auth, no writes, no secrets, no telemetry.

🚀 Install

Using an AI agent? Just tell it: "install the Skillselion MCP from github.com/skillselion/skillselion-mcp" — it can run the commands below itself. For a fully unattended install (no prompts), point it at setup --yes (see Skill autopilot).

Claude Code

claude mcp add skillselion --scope user -- npx -y skillselion-mcp

--scope user makes it available in all your projects (drop it to add to the current project only). Prefer the bleeding edge? Swap skillselion-mcp for github:skillselion/skillselion-mcp to run straight from master.

Claude Desktop / Cursor / Codex

Add this to your MCP config:

{
  "mcpServers": {
    "skillselion": {
      "command": "npx",
      "args": ["-y", "skillselion-mcp"]
    }
  }
}

No build step, no auth, no config. npx fetches a single ~30 KB file. Want full multi-file skill loads without GitHub's 60-req/hr cap? Add a GITHUB_TOKEN to the server's env (raises it to 5000/hr — optional, read-only). See Troubleshooting.

🪄 Skill autopilot (one-command setup)

Want your agent skill-aware automatically, every session?

# Interactive — pick your packs (humans, in a terminal)
npx -y skillselion-mcp setup

# Non-interactive — for agents / CI / scripts (never prompts)
npx -y skillselion-mcp setup --yes
npx -y skillselion-mcp setup --packs frontend,ai-agents --per-pack 5 --auto --yes

It registers the MCP globally (claude mcp add --scope user) and installs a Claude Code SessionStart hook that primes every session with relevant skills and tells the agent to load_skill the right one the moment a task matches — so skills load dynamically, on demand, without you remembering the tools exist.

Choose what it primes you with (instead of a generic global top-10):

Flag	What it does
`--packs <a,b,…>`	one or more packs (or `all`), e.g. `--packs frontend,backend`
`--packs frontend:6,backend:3`	per-pack skill counts
`--per-pack N`	skills per pack (default 3) · `--top N` sets the `popular` count
`--auto`	also adapt each session to the current repo's stack (Next.js → frontend, Python → data, Docker → devops)
`--history`	infer your focus from Claude Code / Codex history (one-time scan at setup)
`--yes` / `--non-interactive`	never prompt (this is the default whenever there's no terminal, so agents never hang)

Packs: popular (default) · frontend · ai-agents · media (generative) · backend · devops · quality · automation.

Safe by design: ~/.claude/settings.json is merged, never clobbered, re-running de-dupes + upgrades the hook in place, it only ever reads the catalog, and the hook runs via an absolute node path so it can't silently fail on PATH. Restart Claude Code to activate.

💬 Example

🗣️ "Set up Postgres for this project."

Instead of winging it, your agent:

search_skillselion({ query: "postgres", type: "skill" }) → finds supabase-postgres-best-practices (244k installs)
load_skill({ id: "skill:supabase/agent-skills#supabase-postgres-best-practices" }) → loads the real SKILL.md into context and drops any bundled scripts/refs into a temp folder
follows that skill — reading its references and running its scripts — like it was installed all along.

With the setup hook above, steps 1–2 happen on their own — the agent loads the skill the moment your task calls for it. ⬆️

🩺 Troubleshooting

The agent never calls load_skill / I don't see the tools. Some clients (including recent Claude Code) surface MCP tools lazily - they don't appear in the initial tool list and are only loaded on demand via a tool search. If your agent doesn't seem to see load_skill, tell it to search its tools for "skillselion" first (e.g. ToolSearch), then call it. The setup hook primes each session with this nudge so it happens on its own.

load_skill returns just the SKILL.md text, without the bundled files. Materializing a skill's full directory uses the GitHub API, which is rate-limited to 60 requests/hr when unauthenticated. Heavy use exhausts that and the server falls back to the raw SKILL.md only. Set a GITHUB_TOKEN (or GH_TOKEN) in the server's environment to raise the limit to 5000/hr and keep full multi-file loads working. It's read-only and optional.

"No skill on Skillselion clearly matches ..." That's the relevance floor working as intended - nothing in the catalog topically matched your task, so loading a merely-popular skill would just waste context. Try a more specific query, or search_skillselion to browse.

Setup said it couldn't auto-register. Run the printed claude mcp add ... line yourself. Auto-registration uses the claude binary; if it isn't on the setup process's PATH (common under npx), setup resolves it from common install locations + a login shell, but a non-standard install can still miss.

🛠 Development

npm install
npm test         # smoke: server starts + exposes the 3 tools
npm run test:live  # hits the live catalog + GitHub (set GITHUB_TOKEN to avoid rate limits)
node index.js    # speaks MCP over stdio

⭐ Like it?

Star this repo — it helps other builders find it.
Browse the full catalog at skillselion.com — search thousands of skills, MCPs & marketplaces, ranked by real installs.
Found a great skill that's missing, or hit a bug? Open an issue — PRs welcome.

📄 License

Built for builders. Find the skill, not the noise. → <a href="https://skillselion.com">skillselion.com</a>