MCP · read-only · no account

Use SpareHive from your AI assistant

The endpoint speaks Streamable HTTP and exposes five read-only tools. It answers from our data set, does not call an LLM for arithmetic, and every number carries an evidence grade.

Endpoint: https://sparehive.com/mcp

Connect

Claude Code (one command)

claude mcp add --transport http sparehive-advisor https://sparehive.com/mcp

Other clients that accept a remote HTTP URL

{
  "mcpServers": {
    "sparehive-advisor": {
      "type": "http",
      "url": "https://sparehive.com/mcp"
    }
  }
}

Verified with Claude Code. Other clients use the same URL; check their documentation for where to paste it.

Five tools

`search_models`	Find a model ID by name or alias, filter by text / image / video.
`model_requirements`	Weights, minimum and recommended VRAM for a quantisation, plus fit verdicts for each GPU.
`fit_check`	Does a model fit a given GPU at a given context? Returns fits / tight / no / unknown, headroom, and alternatives.
`gpu_prices`	Our rental price (RTX 4090 only), competitor median, and market reference per GPU.
`data_status`	Dataset freshness: update dates and whether any file is older than 60 days.

Example

Q: Can an RTX 3080 (10 GB) run Qwen3.8-27B at UD-Q2_K_XL?

A: No, by our formula: about 13.0 GB needed (weights 10.7 + KV + 2 GB overhead) on a 10 GB card. The first alternative is moving to an RTX 4080 16 GB.

What it does not do

No provisioning, no SSH, no billing, no live inventory. We are pre-launch.
No account and no API key. Read-only; it cannot change any data.
Only the RTX 4090 has a published rental price. Other GPUs show 'not priced yet'.
Image and video models have no KV-cache data, so fit verdicts show 'unknown' for them.

Back to the VRAM guide