Vivory MCP Servers

Korean public-data MCP servers for AI agents — KOSIS Statistics Korea (live), BoK ECOS · NEIS · LOCALDATA (planned).

Built and maintained by Vivory. Backend powered by the Vivory Korea Data Gateway.

Why

Korean public statistics (population, GDP, CPI, employment, trade, etc.) are published by official agencies via Korean-only APIs that require local accounts and return JS-literal responses with mixed taxonomies. This repo packages them as Model Context Protocol servers so any MCP-compatible client (Claude Desktop, Claude Code, etc.) can query Korean data in English with zero setup.

Foreign user → Claude → vivory-mcp-* → api.vivory.app → kosis.kr / ecos.bok.or.kr / etc.
                                       (caching, attribution, normalization)

No KOSIS account, no API key, no Korean phone — just install and ask.

Packages

This is a monorepo of MCP server packages. Each is independently installable from PyPI.

Recommended:

• One source only → install the dedicated package (vivory-mcp-kosis)
• All Korean data → install the umbrella (vivory-mcp-korea) — auto-grows as we add sources

Quick install (Claude Desktop)

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "vivory-kosis": {
      "command": "uvx",
      "args": ["vivory-mcp-kosis"]
    }
  }
}

Or to use the umbrella (all Korean data):

{
  "mcpServers": {
    "vivory-korea": {
      "command": "uvx",
      "args": ["vivory-mcp-korea"]
    }
  }
}

Restart Claude Desktop. The tools appear in the tool palette.

PyPI release pending — until then use:

Code

Copy

uvx --from "git+https://github.com/jayjodev/vivory-mcp.git#subdirectory=packages/mcp-server-kosis" vivory-mcp-kosis

Example prompts

Ask Claude (or any MCP client):

"What's Korea's CPI trend over the last 24 months?" "Show me Korean quarterly GDP growth, last 12 quarters." "Search KOSIS for tables related to youth unemployment." "What's the indicator ID for 추계인구?"

Claude picks the right tool automatically.

Architecture

packages/
├── mcp-server-kosis/          # KOSIS standalone MCP server
│   ├── src/vivory_mcp_kosis/
│   │   ├── server.py          # 15 KOSIS tool definitions
│   │   └── client.py          # HTTP wrapper to api.vivory.app
│   ├── pyproject.toml
│   └── README.md
│
└── mcp-server-korea/          # Umbrella MCP server
    ├── src/vivory_mcp_korea/
    │   ├── server.py          # Aggregates tools from all sources
    │   ├── client.py
    │   └── tools/
    │       ├── kosis.py       # KOSIS tool catalog
    │       └── (ecos.py, neis.py, etc. — planned)
    └── ...

All packages are thin wrappers — no API keys, no rate limit handling, no parsing logic. They translate MCP tool calls → HTTP GETs against api.vivory.app, where the heavy lifting lives.

Self-hosting

Set VIVORY_API_BASE to point at your own backend:

export VIVORY_API_BASE="http://localhost:8000/api"

The Vivory backend code is not in this repo (it lives in the private Vivory app monorepo). If you want to run the full stack, you'll need to implement the /api/public-tools/kosis/* endpoints yourself or contact us about licensing.

License

• MCP wrapper code (this repo): MIT
• KOSIS data: 공공누리(KOGL) Type 1 — commercial use permitted with attribution
• Other sources: per-source license (see each package's README)

Every tool response includes an attribution block. When citing or redistributing data, include the attribution_required text.

Contributing

Issues and PRs welcome. Specifically interested in:

• Adding more Korean public data sources to the umbrella
• Improving tool descriptions for better LLM tool-use accuracy
• Adapting for other MCP clients beyond Claude

Project status

• Repo source of truth: this repo (vivory-mcp) — packages here mirror what's published to PyPI
• Vivory backend: private (api.vivory.app)
• Issues / discussions: github.com/jayjodev/vivory-mcp/issues

🇰🇷 Built in Seoul · 🌐 vivory.app