Windlet

GPU workloads by the second

GPU workloads, by the second.

Windlet sells GPU time by size: pick a size from the menu, run a workload, pay for the seconds it ran.

EUR

Each size is a floor a device meets; the hardware behind it is never named. The price is all you pay.

SizeDevice memoryComputeHost memoryDiskSharePer secondPer hourOfferedNow
small Inference of small models, notebooks, short fine-tunes.≥ 8 GiBCUDA 12+—20 GiBwhole€0.00004€0.144YesNow
medium Most open models at 4-bit, LoRA training.≥ 24 GiBCUDA 12.8+≥ 32 GiB50 GiBwhole€0.00011€0.396YesNow
large A whole 32 GiB card: larger models, longer renders.≥ 32 GiBCUDA 12.8+≥ 64 GiB100 GiBwhole€0.00019€0.684YesNow
xlarge A data-centre card with 80 GiB or more.≥ 80 GiBCUDA 12+≥ 128 GiB200 GiBwhole€0.0006€2.16No—

Connect your agent

Any agent with MCP reaches Windlet through the BEST protocol and signs in with the device flow. Add this to its MCP settings, then talk to it.

MCP server 
{
  "mcpServers": {
    "windlet": {
      "command": "npx",
      "args": [
        "-y",
        "@behavioralstate/best-mcp"
      ],
      "env": {
        "BEST_WINDLET_BASE_URL": "https://windlet.io"
      }
    }
  }
}

Endpoint: https://windlet.io