onPanda is an interactive tool for annotating on-policy LLM alignment data and agent trajectories. It uses token-level correction: annotators identify the first inappropriate token, choose a candidate replacement or enter free-form text, then truncate the response and continue generation from the corrected prefix.