In this paper, we propose a novel Transformer-based target tracking framework. In previous Transformer-based trackers, the decoder predicted the spatial location of the target object by learning the query embeddings. However, learned embeddings do not have corresponding physical representations, which makes it impossible to focus on a specific region. In order to make the target query have clear physical meaning, we design the target query in the tracker as the target query based on anchor point. In other words, the anchor is encoded as the target query. In addition, we applied an attention variant RCDA (row-column decoupled attention) that decouples key 2D features into 1D row features and 1D column features and then performs row attention and column attention sequentially. The application of RCDA can achieve better tracking results than basic attention. Our approach is end-to-end and does not require postprocessing steps. Under the same ResNet-50 backbone network, Anchor STARK’s AO score reaches 0.694% on the LaSOT dataset, which is 1.5% higher than STARK-S. Our proposed tracker (Anchor STARK) achieves advanced performance on five challenging short- and long-term benchmarks while satisfying real-time operation. The code and models are available at https://github.com/Renyiam/Anchor-STARK .