Skills MCP

Modular Model Context Protocol host. Skills are folders. Tools from every enabled skill show up on one MCP endpoint.

First skill: Project SEKAI (pjsk_*), backed by the same master JSON sekai.best loads.

Endpoints

Protocol: JSON-RPC 2.0, initialize → tools/list → tools/call. Stateless. CORS open.

Connect a client

Cursor / Claude / any Streamable HTTP client:

{
  "mcpServers": {
    "skills": {
      "url": "https://YOUR-DEPLOYMENT.vercel.app/api/mcp"
    }
  }
}

Add a skill (not related to Sekai)

1. Copy skills/_template to skills/<id>.

2. Export a Skill object (id, title, description, version, tools[]).

3. Each tool needs name, description, JSON Schema inputSchema, and handler(args).

4. Register it in lib/registry.ts:

import { mySkill } from "@/skills/my-skill";

export const skills: Skill[] = [projectSekaiSkill, mySkill];

5. Push to main. Vercel rebuilds.

Keep handlers small and return JSON. The host wraps the result as MCP text content.

Project SEKAI tools

server defaults to jp. Also en, tw, kr, cn.

Local

npm install
npm run dev
# POST http://localhost:3000/api/mcp

curl -s http://localhost:3000/api/mcp \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Deploy

GitHub → Vercel (this repo). Production URL is the MCP base. First pjsk_card call on a cold instance may take a few seconds while cards.json is fetched and slimmed; later calls hit memory + fetch cache.