Loooom MCP: Let Any AI Discover and Install Skills Mid-Conversation

Here's something that bugged me: I built Loooom as a marketplace for Claude Code skills — a place where real people publish reusable AI plugins you can install in seconds. But to actually find a skill, you had to leave your Claude session, visit the site, search around, copy a command, then come back. That's a lot of context-switching for what should be a two-second thing.

So I fixed it. I built @mager/loooom-mcp — an MCP server that lets any AI agent browse and install Loooom skills without ever leaving the conversation.

---

What Loooom Is (Quick Version)

Loooom is a skill marketplace for Claude Code. Skills are essentially plugins — small, focused capabilities you can add to Claude's context. Things like a Japanese language tutor, a git workflow helper, a code review assistant. They're built by humans, free to use, and the whole catalog is on GitHub so you can fork anything you want.

Think of it like npm, but for AI agent behaviors.

---

What MCP Is (Even Quicker Version)

Model Context Protocol is Anthropic's open standard for giving AI agents access to external tools and data. Instead of hardcoding capabilities into a model, MCP lets you plug in servers that expose functions the AI can call — search an API, read a file, hit a database. It's how you extend what an agent can do without retraining anything.

---

What @mager/loooom-mcp Does

The MCP server exposes five tools that map directly onto the Loooom catalog:

list_skills — Pull the full catalog. Every skill, every author. Good for browsing when you're not sure what you're looking for.

search_skills — Find skills by keyword or category. Pass in "japanese" or "git" and get back matching results with descriptions.

get_skill — Deep-dive on a specific skill. Returns the full metadata: what it does, who built it, how to configure it.

list_categories — See what buckets exist in the catalog. Useful when you want to explore a domain without a specific query in mind.

get_install_command — This is the one that closes the loop. Give it a skill name and it hands back the exact Claude Code install command. No copy-paste from a website. Just run it.

Together these five tools turn Loooom from a website you visit into a capability your AI agent has.

---

How to Install

Two ways, depending on your setup.

Claude Desktop

Add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "loooom": {
      "command": "npx",
      "args": ["-y", "@mager/loooom-mcp"]
    }
  }
}

Restart Claude Desktop and the tools are live.

npx (Quick Test)

If you just want to try it:

npx @mager/loooom-mcp

That'll spin up the MCP server locally. Point any MCP-compatible client at it and you're good.

---

A Real Example

Here's what this actually looks like in practice.

You're mid-conversation with Claude, working on something, and you remember you wanted a skill for learning Japanese vocabulary. Old flow: stop what you're doing, go to loooom.xyz, search, find it, copy the install command, come back.

New flow — you just ask:

"Hey Claude, find me a skill for learning Japanese."

Claude calls search_skills with "japanese". Gets back a result: beginner-japanese — a skill that drills vocab and grammar in conversation. Claude surfaces the description, the author, what it does.

You say: "Nice, how do I install that?"

Claude calls get_install_command for beginner-japanese. Hands you back:

claude mcp add beginner-japanese

You run it. Done. Back to what you were doing, but now Claude knows how to teach you Japanese.

That's the whole thing. No tab-switching, no breaking flow. The AI found the tool, the AI told you how to install it, you were never more than a command away.

---

What's Next

A few things I'm working on:

• Publishing to npm — right now you're running it off GitHub. Getting it properly published so the npx install path is clean and fast.
• More skills in the catalog — the MCP server is only as useful as the skills it can find. I'm adding more, and the catalog is open if you want to contribute one.
• Richer metadata — better filtering, tags, ratings. Make search smarter.

The broader idea here is interesting to me: agents should be able to discover and extend their own capabilities dynamically. Today that's "here are five tools to browse a skill catalog." But the direction is toward agents that can identify gaps in what they can do and go find solutions themselves. This is a small step in that direction.

---

If you want to dig in, the code is at github.com/mager/loooom-mcp and the skill catalog is at github.com/mager/loooom. Issues and PRs welcome. And if you build a skill worth adding to Loooom, let me know.