Real-time and energy-constrained systems heavily rely on estimates of the worst-case execution time (WCET) and worst-case energy consumption (WCEC) of code snippets to ensure trustworthy operation. Designing architecture-specific analytical models for time and energy is often challenging and time-consuming. In situations where analytical models are unavailable or incomplete, machine learning (ML) techniques emerge as a promising solution to build WCET/WCEC models. This paper introduces WORTEX, a toolkit for WCET/WCEC estimation of basic blocks based on ML techniques. To ensure the real-world applicability of its models, WORTEX extracts large datasets of basic blocks from real programs and precisely measures their energy consumption/execution time on the physical target platform. The dataset is used to train various WCET/WCEC models using different ML techniques. Experimental results on simple and time-predictable hardware show that even the most basic ML techniques provide accurate results, that never underestimate actual values. We also discuss the use of explainability techniques to gain trustworthiness for the models.