We present MetGen: a sub-daily multi-variable stochastic weather generator implemented as an R library that can be used to perform gap-filling and to extend in time meteorological observation series. MetGen is tailored to provide surrogate series of air temperature, relative air humidity, global radiation and wind speed needed for surface water stress estimation that requires sub-daily resolution. Multiple gauged stations can be used to increase the calibration data although spatial dependence is not modeled. The approach relies on Generalized Linear Models that use, among their covariates, large-scale variables derived from ERA5 reanalyses. MetGen aims at preserving key features of the meteorological variables along with inter-variable dependencies. We illustrate the abilities of MetGen using a case study with three stations in central Tunisia. We consider as alternatives a univariate and a multivariate bias correction techniques along with the un-processed large-scale variables.