MCP Retries and Timeouts

Every remote boundary needs a timeout

Set a total tool budget and shorter downstream timeouts so the server has time to return a useful failure.

Retry only classified failures

Transient network errors and some 5xx responses may be retryable. Validation failures, authorization failures, and most business-rule failures are not.

Use backoff and jitter

Immediate synchronized retries can worsen an outage.

Mutations need idempotency

Before retrying a write, ensure the downstream operation can recognize duplicate attempts.