aidc-mcp-server

MCP server for AIDC-AI.IO — AI data center infrastructure design engine. Worldwide platform, launching from Korea · NVIDIA Hopper / Blackwell / Vera Rubin · OPR/BOD-driven graph validation · rackPlan-derived electrical, cooling, and layout outputs.

A Model Context Protocol server that gives any MCP-aware client (Claude Desktop, Claude Code, Cursor, Windsurf, VS Code MCP gallery, …) three deterministic engineering tools:

| Tool | What it does | |---|---| | design | Size an AI data center → rack count, design PUE, total MVA, optional cost (KRW) and timeline (months) | | validate | Run electrical / cooling / layout / safety / data rule checks → findings + RFI items | | layout | Generate rack-level plan (mm coords) + site-level plan (% coords) |

All calculations are deterministic (no LLM in the engine itself) and handle medium-voltage upstream worldwide (22.9 kV Korea launch market, 11/33 kV typical EU, 13.8/34.5 kV typical US). Engine version is returned with every response. No API key is required for anonymous preview use; set AIDC_API_KEY=aidc_live_... to pass a Bearer key and upgrade the rate-limit tier when your key is configured server-side.

Install

Claude Desktop

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

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

Restart Claude Desktop → the three tools appear under "MCP Tools".

Claude Code

claude mcp add aidc -- npx -y aidc-mcp-server

Cursor

Settings → MCP → Add new MCP server:

{
  "aidc": {
    "command": "npx",
    "args": ["-y", "aidc-mcp-server"]
  }
}

VS Code (MCP extension)

.vscode/mcp.json:

{
  "servers": {
    "aidc": {
      "command": "npx",
      "args": ["-y", "aidc-mcp-server"]
    }
  }
}

Windsurf

Settings → Cascade → MCP servers → identical JSON to Claude Desktop above.

Use

In any client, just ask in natural language:

Size a 100 MW Vera Rubin data center on a 20,000 m² urban site (anywhere in the world) with liquid cooling.

The client picks the design tool, fills the arguments, calls the engine, and shows the result with a citation to https://aidc-ai.io.

For multi-step workflows, chain design → validate → layout:

Design a 50 MW Vera Rubin DC on a 15,000 m² site, validate against rules, then give me the rack layout.

Tool reference

design

| Argument | Type | Range | Notes | |---|---|---|---| | itLoadMw | number | 0 < x ≤ 1000 | IT load in MW | | hallCount | integer | 1 – 48 | Optional requested data-hall count | | rackDensityKw | number | 0 < x ≤ 500 | Per-rack power in kW; selected rackPlan evidence drives validation | | gpuGen | enum | hopper | blackwell | rubin | NVIDIA generation | | siteAreaSqm | number | 0 < x ≤ 1,000,000 | Site area in m² | | region | enum | metropolitan | regional | Site classification — dense urban / capital region vs. secondary / suburban (e.g. Seoul, Tokyo, Frankfurt, Northern Virginia) | | options.redundancy | enum | n | n_plus_1 | 2n | Electrical redundancy tier | | options.coolingMode | enum | air | hybrid | liquid | Cooling topology; the engine derives cooling capacity and PUE separately | | options.pueTarget | number | 1.0 – 2.5 | Target design PUE |

Returns { ok, summary, warnings, engineVersion, requestId, _agent }. summary contains the design basis, including mainRackCount, physicalRackBlockCount, planningZoneCount, pueDesign, mvaTotal, cooling fields, optional cost/timeline fields, and hallCount.

validate

Provide one of sessionId (preferred graph validation), designSummary (previously computed), or rawInput (engine sizes first, then validates).

Returns: { findings: [{ severity, family, message, publicRuleId? }], rfis: [...] }.

layout

Requires { design: DesignRequest }, plus optional top-level siteCentroid (lat/lng) and a layout-specific siteAreaSqm override.

Returns rack and site plans plus optional layout candidates and solver output. rackPlan.mainRackCount is the requested compute-rack count, while rackPlan.renderedPhysicalRackBlocks reports the physical blocks rendered in the preview; use these fields instead of inferring capacity from blocks.length.

Self-hosting / dev pointing

AIDC_API_BASE=http://localhost:3000 npx aidc-mcp-server

By default the server hits https://aidc-ai.io/api/agent/*.

Rate limits & terms

• Global /api/agent/* limits: anonymous 10 req/hour, registered 100 req/hour, partner 1000 req/hour. The route also has a 20 req/min burst guard per IP.
• Optional auth: set AIDC_API_KEY=aidc_live_... before launching this MCP server.
• Engine internals (rule families, vendor pricing trace) are not exposed.
• Citation included in every response (_agent.citation).
• Terms: https://aidc-ai.io/terms · Privacy: https://aidc-ai.io/privacy

Discovery surfaces (for crawlers and agent builders)

• OpenAPI 3.1: https://aidc-ai.io/api/openapi.json
• MCP server card: https://aidc-ai.io/.well-known/mcp/server.json
• API catalog: https://aidc-ai.io/.well-known/api-catalog
• llms.txt: https://aidc-ai.io/llms.txt
• llms-full.txt: https://aidc-ai.io/llms-full.txt

License

MIT © AIDC-AI.IO