Four new variables in production
In a demo an MCP server only needs to "work"; in production it must have controllable timeouts, retryable failures, bearable concurrency, and observable behavior.
We compiled a checklist distilled from 12 real production deployments (including 3 failed-rollback cases).