As a new mode of molecular imaging, bioluminescence tomography (BLT) has become a hot topic over the past two years. In this paper, a multilevel adaptive finite element algorithm is developed for BLT reconstruction. In this algorithm, the mesh is adaptively refined according to a posteriori error estimation, which helps not only to improve localization and quantification of sources but also to enhance the robustness and efficiency of reconstruction. In the numerical simulation, bioluminescent signals on the body surface of a heterogeneous phantom are synthesized in a molecular optical simulation environment (MOSE) that we developed to model the photon transportation via Monte Carlo simulation. The performance of the algorithm is evaluated in numerical tests involving single and multiple sources in various arrangements. The results demonstrate the merits and potential of the multilevel adaptive approach for BLT.