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.
Use this checklist as a starting point and adapt thresholds to your workload, client behavior and recovery requirements.
Stability config
Timeouts: the MCP protocol doesn't enforce timeouts, so you must configure them client-side. We suggest a 30s default for tool calls, with long tasks declared explicitly…
Retries: idempotent read-only tools can retry safely; write operations must implement a dedup key…
Observability plan
A complete log-collection plan: which fields must be recorded (tool name, latency, token cost, error type), how to feed it into your existing APM, and our open-source JSON log schema.