CoTracker is a Transformer-based model capable of joint tracking of dense points in video sequences. Unlike most existing state-of-the-art methods, which track points independently and neglect their correlations, CoTracker demonstrates that joint tracking can significantly improve tracking accuracy and robustness. We have introduced several innovative techniques, including the concept of virtual trajectories, enabling CoTracker to jointly track up to 70,000 points. Furthermore, CoTracker operates causally over short time windows, making it suitable for online tasks, yet through training over longer video sequences, this significantly improves long-term tracking. We exhibit compelling tracking results, where points can be tracked for extended periods even during occlusions or when they leave the field of view. Quantitatively, CoTracker outperforms all recent trackers on standard benchmark tests, with substantial advantages in most cases.