The reconstruction of trees from point clouds that were acquired with terrestrial LiDAR scanning (TLS) may become a significant breakthrough in the study and modelling of tree development. Here, we develop an efficient method and a tool based on extensive modifications to the skeletal extraction method that was first introduced by Verroust and Lazarus in 2000. PypeTree, a user-friendly and open-source visual modelling environment, incorporates a number of improvements into the original skeletal extraction technique, making it better adapted to tackle the challenge of tree perennial tissue reconstruction. Within PypeTree, we also introduce the idea of using semi-supervised adjustment tools to address methodological challenges that are associated with imperfect point cloud datasets and which further improve reconstruction accuracy. The performance of these automatic and semi-supervised approaches was tested with the help of synthetic models and subsequently validated on real trees. Accuracy of automatic reconstruction greatly varied in terms of axis detection because small (length < 3.5 cm) branches were difficult to detect. However, as small branches account for little in terms of total skeleton length, mean reconstruction error for cumulated skeleton length only reached 5.1% and 1.8% with automatic or semi-supervised reconstruction, respectively. In some cases, using the supervised tools, a perfect reconstruction of the perennial tissue could be achieved.