Abstract Objective. Neurofeedback (NF) is a cognitive training procedure based on real-time feedback (FB) of a participant’s brain activity that they must learn to self-regulate. A classical visual FB delivered in a NF task is a filling gauge reflecting a measure of brain activity. This abstract visual FB is not transparently linked—from the subject's perspective—to the task performed (e.g., motor imagery). This may decrease the sense of agency, that is, the participants’ reported control over FB. Here, we assessed the influence of FB transparency on NF performance and the role of agency in this relationship. Approach. Participants performed a NF task using motor imagery to regulate brain activity measured using electroencephalography. In separate blocks, participants experienced three different conditions designed to vary transparency: FB was presented as either 1) a swinging pendulum, 2) a clenching virtual hand, 3) a clenching virtual hand combined with a motor illusion induced by tendon vibration. We measured self-reported agency and user experience after each NF block. Main results. We found that FB transparency influences NF performance. Transparent visual FB provided by the virtual hand resulted in significantly better NF performance than the abstract FB of the pendulum. Surprisingly, adding a motor illusion to the virtual hand significantly decreased performance relative to the virtual hand alone. When introduced in incremental linear mixed effect models, self-reported agency was significantly associated with NF performance and it captured the variance related to the effect of FB transparency on NF performance. Significance. Our results highlight the relevance of transparent FB in relation to the sense of agency. This is likely an important consideration in designing FB to improve NF performance and learning outcomes.