Machine translation (MT) enables a group of people who do not share a common language to work together as a team. Previous studies have investigated the characteristics of MT-mediated communication in laboratory settings and suggested various ways to improve it. Yet, few studies have investigated how MT is actually used outside the lab. We still lack an understanding of how MT is used in real-world settings, particularly when people use it in face-to-face situations. In this paper, we report on an ethnographic study of a multilingual children workshop using MT to communicate with each other in real world. We studied how children use various communication methods such as gesture and internet to compensate for the mistranslations of MT. For example, children tried to understand poorly translated messages by reading the alternative translations and used web browsers to search for pictures of unknown objects. Finally, we propose design implementations based on our findings.