What is an MCP Transport?

MCP deliberately separates its message format, JSON-RPC 2.0, defining requests, responses, and notifications, from how those messages physically move between processes. This separation lets the same protocol semantics work whether a server is a local subprocess on the same machine or a remote service reached over the internet — the client and server agree on a transport at connection time, but the tool-calling, resource-reading, and prompt logic above it stays identical either way. The two transports in current use serve different deployment shapes: stdio is used when the host itself launches the server as a child process, giving it direct, low-latency access with no network hop, appropriate for local-machine capabilities. Streamable HTTP is used when the server runs independently, potentially shared across many clients or hosted in the cloud, and needs standard web infrastructure like TLS, load balancers, and authentication headers to work safely at scale. Choosing the right transport is mostly a deployment decision rather than a protocol-capability one: a filesystem server almost always uses stdio, while a SaaS-integration server almost always uses Streamable HTTP, because the resource each wraps naturally lives on one side of that machine boundary or the other.