Nankai Trough MCP server

nankai-trough-mcp

Nankai Trough (南海トラフ地震) earthquake hazard and building-safety engine for AI assistants. Ask "how bad is the Nankai Trough quake, how far does it reach, and what does that mean for a house built in 1978?" and get official Japanese government figures, every one cited and dated, with a route to the official maps for your exact address. It never tells you you're "safe."

It reports what 内閣府 (Cabinet Office), 地震本部, 気象庁, and 国土交通省 actually publish, and where a figure isn't verifiable, it points you at the source instead of inventing one. Bilingual EN/JA. Runs locally, no API keys.

Live companion map: https://nankaitrough.bymarsel.me (this is the open data engine behind it, the same official data as an explorable map).

⚠️ Not a prediction, not a verdict. This explains official data and routes you to official guidance and a professional seismic diagnosis (耐震診断). Always confirm with your municipality and official hazard maps. In an emergency, follow official evacuation instructions.

Why it exists

Most people underestimate the reach. The 30-year probability was revised by 地震本部 in September 2025 away from the old single "80%" to a two-model range (60–90%+ (higher model) / 20–50%). Intensity-7 shaking is projected across 10 prefectures, and 764 municipalities in 31 prefectures face major shaking or a 3 m+ tsunami. "I'm inland, so I'm fine" is exactly the assumption this corrects.

The authoritative data is real, but it's scattered across five agencies and mostly buried in PDFs. This server makes it cited and usable from an AI assistant, without ever inventing a number.

Data rules

• No verdicts. It reports official figures + plain-language meaning, never "your home is safe/unsafe."
• Every figure cites its official source + date. If a value isn't verifiable in a primary source, it points to the source instead of guessing.
• Probabilistic ≠ scenario. J-SHIS (all-source probability) is never presented as the Nankai scenario (Cabinet Office).
• Building safety can't be looked up. It's classified from the build year + structure you provide. No fabricated per-building data.

Install

claude mcp add nankai -- npx -y nankai-trough-mcp

JSON config

The server config is the same across clients; only the file path differs.

Claude Code: ~/.claude/.mcp.json

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

Cursor: ~/.cursor/mcp.json

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

VS Code (GitHub Copilot): .vscode/mcp.json

{
  "servers": {
    "nankai": { "type": "stdio", "command": "npx", "args": ["-y", "nankai-trough-mcp"] }
  }
}

Claude Desktop: claudedesktopconfig.json

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

Config path:

• macOS: ~/Library/Application Support/Claude/claudedesktopconfig.json
• Windows: %APPDATA%\Claude\claudedesktopconfig.json

From source

git clone https://github.com/mrslbt/nankai-trough-mcp.git
cd nankai-trough-mcp
npm install && npm run build

Point the client config at the built entry:

{ "command": "node", "args": ["/absolute/path/to/nankai-trough-mcp/dist/index.js"] }

Tools

Shortened. The full README is on GitHub.