Multivariate time series are vital for capturing complex interactions among variables in domains like finance, e-commerce, and climate science. However, existing research has largely focused on custom univariate models, leaving gaps in multivariate scenarios and foundation models capable of universal forecasting. We address these gaps with two contributions: a novel framework that adapts traditional tokenization techniques to multivariate time series, integrating multiple target and feature series into a unified model allowing us to leverage existing foundation models for language; and an innovative synthetic data generation process to overcome data scarcity, enabling robust model training. Our approach handles diverse covariates and is validated through extensive experiments, demonstrating superior performance over current state-of-the-art methods.