How does Streamable HTTP work as an MCP transport and when should you choose it over stdio?

**Streamable HTTP** is how a remote MCP server and its host talk over the network. The whole thing lives behind one URL. The client sends a JSON-RPC message with an HTTP POST, and the server answers. The clever part is the response. For a quick reply, the server just returns a normal JSON body and the request ends. For a slow tool, or when the server needs to push notifications, it upgrades the same response into a stream using Server-Sent Events. So one endpoint covers both fast and slow cases. Each conversation gets a session id, sent in a header. That lets the server keep per-session state and lets the client reconnect after a dropped connection without starting over. You choose Streamable HTTP whenever the server is remote: a cloud VM, a serverless function, or a managed service. You use stdio instead when the server runs as a local child process on the same machine as the host, since stdio needs no network at all.

Streamable HTTP is the current remote transport for MCP. It replaced the older HTTP plus SSE design. Everything runs through a single HTTP endpoint. The client sends each JSON-RPC message as an HTTP POST. The server inspects the request and picks one of two response modes. For a short result it returns a normal JSON body and closes the request. For a long-running call, or to deliver server-initiated notifications, it sets the response to an SSE stream and emits events over that open body. Sessions are tracked with a header. On initialize, the server may return an `Mcp-Session-Id`, and the client echoes it on every later request. Because streamed events carry IDs, a client that drops can reconnect and resume from the last event it saw, so it never misses messages. Choose Streamable HTTP whenever the server is remote or deployed: a cloud host, a Cloudflare Worker, or any managed service reachable over the network. It works on ordinary HTTP infrastructure, including serverless. Reach for stdio only when the server is a local child process on the same machine, where a network hop would add pure overhead.