The goal of ConTextNet is to discover influential text features from a corpus given an outcome of interest, according to the methodology presented in Ayers et al. (2024). These text features can be used to estimate causal effects on the outcome (Fong & Grimmer 2016, 2021), or can be viewed as exploratory evidence to guide confirmatory analyses where researchers design a small number of treatment texts. To do this, ConTextNet walks users through building, training, tuning, and interpreting a neural network with convolutional layers. ConTextNet relies heavily on R Keras for modeling, and will likely require users to run at least some operations on a high performance computing cluster (HPC) for corpora larger than a few thousand documents. The package will provide documentation and helper functions to guide users through the necessary Python dependency installations and to streamline interactions with common HPC interfaces.