Runpod MCP Server

npx -y @runpod/mcp-server@latest

{
  "mcpServers": {
    "runpod-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "@runpod/mcp-server@latest"
      ]
    }
  }
}

Runpod rents GPUs by the minute, and this server puts the whole control plane behind tool calls: pods, serverless endpoints, templates and network volumes. Two connection modes exist and the difference is the interesting part. The hosted endpoint at mcp.getrunpod.io signs the user in through Runpod OAuth and keeps no key on disk, while the local npx process reads RUNPOD_API_KEY out of the client config. Either way each request carries the caller's own token, which is forwarded to the Runpod API and never persisted server-side, so one hosted instance never mixes credentials between users. One migration detail catches people out: on the v2 REST API serverless endpoints are image-based, so you pass an image plus a GPU pool value from list-gpu-types instead of a template id.