Building extremely agile drones that can comfortably navigate indoor environments requires solving a number of challenges in hardware as well as planning and perception algorithms. We'll present a new system that allows the development of vision-based agile drones. We'll present a micro-aerial vehicle that is based on the NVIDIA Jetson TX2 computer and can fly in complex indoor environments using visual and inertial measurements only. We'll also present a novel development environment that allows us to field a wide range of planning and perception algorithms, including visual simultaneous localization and mapping, differential pursuit-evasion games, and deep reinforcement learning. This work is partly supported by the MIT NVAIL.