Despite the technological advancements, the employment of passive brain computer interface (BCI) out of the laboratory context is still challenging. This is largely due to methodological reasons. On the one hand, machine learning methods have shown their potential in maximizing performance for user mental states classification. On the other hand, the issues related to the necessary and frequent calibration of algorithms and to the temporal resolution of the measurement (i.e. how long it takes to have a reliable state measure) are still unsolved. This work explores the performances of a passive BCI system for mental effort monitoring consisting of three frontal electroencephalographic (EEG) channels. In particular, three calibration approaches have been tested: an intra-subject approach, a cross-subject approach, and a free-calibration procedure based on the simple average of theta activity over the three employed channels. A Random Forest model has been employed in the first two cases. The results obtained during multi-tasking have shown that the cross-subject approach allows the classification of low and high mental effort with an AUC higher than 0.9, with a related time resolution of 45 seconds. Moreover, these performances are not significantly different from the intra-subject approach although they are significantly higher than the calibration-free approach. In conclusion, these results suggest that a light (three EEG channels) passive BCI system based on a Random Forest algorithm and cross-subject calibration could be a simple and reliable tool for out-of-the-lab employment.