Description: The Amazon forests are one of the largest ecosystem carbon pools on Earth. Although more frequent and prolonged future droughts have been predicted, the impacts have remained largely uncertain, as most land surface models (LSMs) fail to capture the vegetation drought responses. In this study, the ability of the LSM JSBACH to simulate the drought responses of leaf area index (LAI) and leaf litter production in the Amazon forests is evaluated against artificial drought experiments. Based on the evaluation, improvements are implemented, including a dependency of leaf growth on leaf carbon allocation and a better representation of drought-dependent leaf shedding. The modified JSBACH is shown to capture the drought responses at two sites and across different regions of the basin. It is then coupled with an atmospheric model to simulate the carbon and biogeophysical feedbacks of drought under future climate. We separate the drought impacts into (a) the direct effect, resulting from drier soil and stomatal closure, which does not involve a change in canopy structure, and (b) the LAI effect, resulting from leaf shedding and involving canopy response. We show that the latter accounts for 35% of reduced land carbon uptake (9 ± 10 vs. 26 ± 7 g/m2/yr; mean ± 1 sd) and 12% of surface warming (0.09 ± 0.03 vs. 0.7 ± 0.07 K) during the late 21st century. A north-south dipole of precipitation change is found, which is largely attributable to the direct effect. The results highlight the importance of incorporating drought deciduousness of tropical rainforests in LSMs to better simulate land-atmosphere interactions in the future.