MCP server

What it is for

A language model asked "what does an H100 cost per hour" answers from whatever it memorised months ago. Connected to this server, it calls get_gpu_price and answers with today's number, the market median behind it and the page it came from. The same applies to sizing questions ("which GPU runs a 70B model") and to budgets ("what does a fine-tuning weekend cost").

The server is deliberately narrow: it reads the public catalogue and the public inventory. It cannot see an account, cannot deploy, cannot spend. Deployment stays in the console and the authenticated REST API.

Connect a client

Transport is Streamable HTTP: JSON-RPC 2.0 over POST to a single URL. Most clients only need the URL.

{
  "mcpServers": {
    "powergpu": {
      "type": "http",
      "url": "https://powergpu.ai/mcp"
    }
  }
}

Claude Code takes it in one command; other CLIs are similar.

claude mcp add --transport http powergpu https://powergpu.ai/mcp

For a client that only speaks stdio, bridge it:

npx -y mcp-remote https://powergpu.ai/mcp

Tools

Try it without a client

It is an ordinary HTTP endpoint, so curl works:

# what tools exist
curl -s -X POST https://powergpu.ai/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | jq '.result.tools[].name'

# today's H100 SXM price
curl -s -X POST https://powergpu.ai/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call",
       "params":{"name":"get_gpu_price","arguments":{"gpu":"h100-sxm"}}}' \
  | jq -r '.result.content[0].text'

# cost of 8x H100 for a 40-hour training run, with 2 TB of storage
curl -s -X POST https://powergpu.ai/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":3,"method":"tools/call",
       "params":{"name":"estimate_cost","arguments":{"gpu":"h100-sxm","hours":40,
                  "num_gpus":8,"storage_gb":2000}}}' | jq -r '.result.content[0].text'

A GET on the same URL returns a plain description of the server, which is handy when you are wiring things up.

Response shape

Every tool returns human-readable text plus a structuredContent object, so a model can quote the prose and a program can read the fields.

{
  "result": {
    "content": [{ "type": "text", "text": "NVIDIA H100 SXM (80 GB): $1.428/hr on-demand …" }],
    "structuredContent": {
      "slug": "h100-sxm",
      "vram_gb": 80,
      "usd_per_gpu_hour": { "on_demand": 1.428, "interruptible": 0.714, "reserved": 0.928 },
      "market_median_usd_hr": 2.0414,
      "percent_below_median": 30,
      "url": "https://powergpu.ai/gpu/h100-sxm"
    }
  }
}

Limits and guarantees

• Read-only. No tool writes anything. There is nothing to authorise and nothing to revoke.
• No account data. The server has no notion of users, balances or instances; it sees the public catalogue only.
• No key, no rate limit worth worrying about. Ordinary abuse protection applies at the edge, nothing more.
• Live prices. Every figure comes from the same sheet the website renders, at the market snapshot in force (2026-09-14).
• Protocol. MCP 2025-06-18 over Streamable HTTP. Batched JSON-RPC requests are accepted.

Other machine-readable surfaces

• /llms.txt — the site index with the key facts, for assistants that read it.
• /llms-full.txt — every page of the site converted to Markdown, one file.
• Markdown for any page — append .md to a page URL (/pricing.md) or send Accept: text/markdown.
• /openapi.json — OpenAPI 3.1 description of the REST API, for function calling and client generation.
• /gpu-price-index.json and .csv — the dated price dataset, CC BY 4.0.