MCP Server Cold Start

An agent using a Python MCP search server on Cloud Run saw 3,800ms first-call latency for 40% of sessions (those starting after the 10-minute idle scale-down). Moving to an always-on $6/mo VPS eliminated cold starts and reduced average first-call latency from 1,700ms to 380ms.

MCP Server Cold Start is relevant to google. Scavio provides a unified API to access data from all of these platforms.

Cold start latency varies significantly by deployment model. Self-hosted MCP servers running as serverless functions (AWS Lambda, Vercel Functions, Google Cloud Run) scale to zero after a configurable idle period (typically 5-15 minutes). Cold start for a Node.js MCP function is 800-2,000ms; Python is 1,500-4,000ms due to import overhead. A Docker container cold start on Cloud Run is 2,000-6,000ms depending on image size. Always-on deployments (VPS, dedicated container, ECS with minimum 1 task) eliminate cold starts entirely at the cost of idle compute. A $6/month VPS running a Node.js MCP server keeps the process warm indefinitely — cheaper than the engineering cost of debugging cold start failures in production. Hosted MCP endpoints provided by API vendors (including MCP-compatible search APIs) are always-on by design; cold starts are the vendor's problem, not the developer's. For agent workflows where search is called multiple times per session, a 2-4 second cold start on the first call is tolerable. For workflows where search is called once per session, the cold start represents a large fraction of total session time and should be mitigated with a keep-alive ping (lightweight OPTIONS request every 5 minutes).