Machine-learning potentials (MLIPs) have been a breakthrough for computational physics in bringing the accuracy of quantum mechanics to atomistic modeling. To achieve near-quantum accuracy, it is necessary that neighborhoods contained in the training set are rather close to the ones encountered during a simulation. Yet, constructing a single training set that works well for all applications is, and likely will remain, infeasible, so, one strategy is to supplement training protocols for MLIPs with additional learning methods, such as active learning, or fine-tuning. This strategy, however, yields very complex training protocols that are difficult to implement efficiently, and cumbersome to interpret, analyze, and reproduce. To address the above difficulties, we propose AutoPot, a software for automating the construction and archiving of MLIPs. AutoPot is based on BlackDynamite, a software operating parametric tasks, e.g., running simulations, or single-point ab initio calculations, in a highly-parallelized fashion, and Motoko, an event-based workflow manager for orchestrating interactions between the tasks. The initial version of AutoPot supports selection of training configurations from large training candidate sets, and on-the-fly selection from molecular dynamics simulations, using Moment Tensor Potentials as implemented in MLIP-2, and single-point calculations of the selected training configurations using VASP. Another strength of AutoPot is its flexibility: BlackDynamite tasks and orchestrators are Python functions to which own existing code can be easily added and manipulated without writing complex parsers. Therefore, it will be straightforward to add other MLIP and ab initio codes, and manipulate the Motoko orchestrators to implement other training protocols.