psychosis-guard

psychosis-guard is an open-source, model-agnostic safety middleware designed for long-running chatbot conversations using large language models. It detects and intervenes in slowly escalating delusional or harmful patterns across a conversation trajectory that typical turn-local safety filters miss. It runs as an HTTP proxy or Python library, supports multiple LLM backends, and provides graduated, clinically-informed interventions to improve chatbot safety over time.