An adaptive multiscale finite element method is proposed for solving the mechanical problems of heterogeneous material in elasticity with aim to improve the computational accuracy as needed and decrease the computational resources as much as possible.In this method, the computational domain is divided into several Unit Cells (UCs) with the specific size, and the initial Numerical Base Functions (NBFs) are constructed by the linear boundary conditions for each UC.The NBFs have been proved to satisfy the rigid displacement properties previously and act as a bridge between the macroscopic and microscopic scales.We have proved that the suitable number of UCs would minimize the computational quantity for the decomposition of coefficient matrix; however, such results may not meet the requirement of accuracy at some special locations of the structure, such as the zone of stress concentration or strain localization.In our method, the computational accuracy is improved by adjusting the node density dynamically on the interfaces of the UCs according to the posteriori error estimator, which is based on the superconvergent patch recovery (SPR) for stresses proposed by Zienkiewicz and Zhu.Only the patches close to the interfaces are taken into consideration for the stress recovery and the hierarchical NBFs corresponding to the new generated nodes are updated independently without affecting the previous NBFs.Thus, the original model needs not to be remeshed and the decompositions of the coefficient matrix in both macroscopic and microscopic are implemented only at the first step of calculations.Therefore, the accuracy of the model could be improved without increasing too much computational costs in the method.In addition, the parallel computation could be implemented easily as the construction of NBFs is independent for each UC.Several examples are presented to demonstrate the accuracy and efficiency of the method.