One of the biggest problems facing online platforms today is the prevalence of so-called "toxic" behavior, such as personal attacks, harassment, and general incivility.While a common computational approach for addressing this problem has been developing algorithms to detect toxicity, we argue that this approach reflects an overly narrow view of online community governance, catering specifically to the use case of platform-driven, centralized content moderation, while overlooking an equally important perspective: that of the communities of ordinary users who interact on these platforms. Therefore, this dissertation takes on the following question: how can technology support members of online communities in having healthier interactions, and thereby proactively prevent toxicity from taking root? We take a combined social and technical approach to answering this question.From the social perspective, we begin with a close examination of existing practices of online community governance: drawing from literature in diverse fields ranging from computer science to sociology, law, and political science, we identify concrete ways in which online communities proactively prevent toxicity and promote pro-social norms, and conduct interviews to gain more qualitative insights. These insights guide our technical approach: inspired by interview participants' explanations of how they can intuitively tell whether a conversation might later derail into toxicity, we formalize such derailment forecasting as a novel computational task and argue that solving it requires a new class of conversational forecasting models. Finally, bringing together the technical and social aspects, we develop a first-of-its-kind concrete implementation of a conversational forecasting model and evaluate it via an "in-the-wild" user study involving ordinary users in a real online community. We conclude by looking back on our findings thus far and comparing them with our higher-level, long-term goals for this work.From this comparison, we identify current shortcomings and unanswered questions that should be tackled in future work, and pull in insights from recent developments in machine learning, natural language processing, and computational social science to build a concrete roadmap of next steps.