Providing reliable cancer information with a large language model-based agent is challenging due to domain nuance, clinical safety constraints, and the risk of hallucinations. In oncology, standard retrieval augmented generation (RAG) lacks triage and medical verification. We present CancerRAGent, a safety-aware RAG for oncology question answering. On a curated oncology PubMed subset, it combines hybrid retrieval, reranking, and evidence-grounded generation with red-alert triage to deliver a concise answer, safety-guided exploration, and a suggested next question. Users receive concise, cancer-focused answers with a clear follow-up question and guided exploration—features absent in vanilla RAG. This demo targets patients and caregivers seeking evidence-based cancer information. The system does not provide diagnoses or personalized treatment plans. Our code and demo are here: https://github.com/oncologyxai/CancerRAGent.