GPUFind 推出 MCP 服务，在编码智能体中查询 56 家云厂商 GPU 价格

MCP server

Cloud GPU prices, inside your coding agent.

Ask Claude Code, Codex or Cursor where a GPU is cheapest and get an answer from 56 providers' published prices — each with how recently we confirmed it and a link to the provider's own pricing page, so you can check it yourself.

500 free tool calls a month · sign in with Google · no card

You askWhat's the cheapest on-demand H100 SXM right now?

GPUFind returns

GPU.ai$1.49/GPU-hrconfirmed 41m ago · source gpu.ai
Lium$1.85/GPU-hrconfirmed 42m ago · source lium.io
RunPod Community Cloud$2.69/GPU-hrconfirmed 3h ago · source runpod.io

Today's real data, the same numbers the tool returns. Your agent gets them as structured results and can rank, filter or cost a run with them.

Set it up

Sign in with Google at gpufind.cloud/login.
Create an API key on your account page. It is shown once — copy it then.
Add GPUFind to your client with one of the snippets below, using your key in place of gpf_live_….

Claude Code

claude mcp add --transport http gpufind https://www.gpufind.cloud/mcp \
  --header "Authorization: Bearer gpf_live_…"

Codex

In ~/.codex/config.toml, then export GPUFIND_API_KEY=gpf_live_… in your shell:

[mcp_servers.gpufind]
url = "https://www.gpufind.cloud/mcp"
bearer_token_env_var = "GPUFIND_API_KEY"

Cursor

In ~/.cursor/mcp.json (or .cursor/mcp.json for one project), with GPUFIND_API_KEY set in your environment:

{
  "mcpServers": {
    "gpufind": {
      "url": "https://www.gpufind.cloud/mcp",
      "headers": { "Authorization": "Bearer ${env:GPUFIND_API_KEY}" }
    }
  }
}

Any other MCP client that supports remote servers works the same way: the URL https://www.gpufind.cloud/mcp and the header Authorization: Bearer <your key>.

What people use it for

Find where to run a job, right now

“I need 8× H100 SXM on demand today. Who's cheapest, and is it in stock?”

cheapest_offers with gpu_count 8 and in_stock_only ranks the offers that fit, with stock where the provider publishes it and a link to rent each one.

Cost a run before you start it

“Fine-tuning this model needs about 16 A100 80GB hours. What does that cost at the three cheapest providers?”

Your agent pulls live per-GPU-hour rates and does the arithmetic in your conversation — a budget built on today's published prices instead of a number someone remembered.

Pick the card, not just the provider

“My model needs 48 GB of VRAM. Which cards have that, and which is cheapest per hour?”

search_gpus finds every tracked card matching the memory you need, with the cheapest on-demand rate for each, so you compare L40S, A6000 and RTX 6000 Ada side by side.

Sanity-check a quote

“A reseller quoted us $2.90/hr for an H100. Is that a good price?”

get_gpu_prices returns every published offer for the card across providers, so your agent can tell you where the quote sits — and which listed prices beat it.

Check where a number came from

“Where does that cheapest price come from? Show me the source.”

get_offer_provenance returns the exact text the provider's page said, the URL it was read from, when, the SHA-256 of our archived copy, and every step applied to reach the figure.

The tools

search_gpus: Find GPU models by name or memory, with how many providers rent each and its cheapest on-demand rate.
cheapest_offers: The cheapest on-demand per-GPU offers for one card, ranked — optionally by card count and in stock only.
get_gpu_prices: Every current offer for one card from every provider, grouped by billing type: on-demand, spot, committed, reserved.
list_providers: Every provider we track: how many cards and offers, the cheapest rate, and what kind of provider it is.
get_offer_provenance: The audit trail behind one price: the source text, URL, time, archive hash and every transformation.

Why you can trust the answers

Published list prices, read from each provider's own pricing page every one to six hours, and archived byte for byte. Nothing negotiated, committed or enterprise.
Every price says when we last confirmed it and links to the page it came from. A source that stops answering is flagged stale, so an old number never reads as current.
Only like is ranked against like: on-demand per-GPU prices are compared with each other. Spot, committed and per-node prices are listed, labelled, in their own groups.
“Contact us” stays “contact us”. A price a provider doesn't publish comes back as not published — never as zero or a guess.
Nothing is paid placement. No provider can pay to rank, and your usage never affects ranking.

Limits

500 tool calls a month, free, shared by all your keys and reset on the 1st (UTC). Connecting and listing tools don't count. Past 400, results include a note your agent will pass on; at 500 it tells you the date it resets.
Up to 30 requests a minute, and up to three active keys. A leaked key can be rotated in one click.
Works today with clients that take an API key — Claude Code, Codex, Cursor and others. Connecting from the claude.ai or ChatGPT apps needs a sign-in flow we haven't built yet.