Building Production MCP Servers: What the Spec Won’t Tell You

Scroll through developer forums this month and you’ll see the same obituary in every thread: MCP is dead. Eric Holmes’ “MCP is dead, long live the CLI” hit the top of Hacker News. Pieter Levels called MCP “just as useless of an idea as LLMs.txt.” Thoughtworks put “naive API-to-MCP conversion” in the Hold ring of their Tech Radar.

I’ve been building on MCP since the spec dropped. The obituary reads the symptoms right and the mechanism wrong.

Here’s the problem. MCP isn’t a protocol — it’s a cost model wearing a protocol’s clothes. JSON-RPC is just the envelope. What MCP actually does is delegate three expensive problems to three different places without naming any of them: token cost to the model, tool-selection cost to the model’s attention, and sanitization cost to you. When your server works in demo and collapses in production, it’s not the protocol failing. It’s the cost model arriving all at once.

This article shows you what the spec hands off without saying so. By the end, you’ll know which cost goes where, what the spec leaves for you to…