You should go for differencing. Without it, you would have I(1) variables that are not cointegrated on the left and the right hand sides of each equation, causing one side to diverge from the other unless (1) the slope coefficients of own lags sum to one and (2) the slope coefficients of lags of other variables sum to zero for each variable (assuming there is no cointegration between subgroups of variables). E.g. take one equation of a bivariate VAR(1) model for variables $(y,x)$:
$$
y_{t}=\beta_0+\beta_1 y_{t-1}+\beta_2 x_{t-1}+\varepsilon_t
$$
where $y_t$ and $x_t$ do not share a common trend and thus are diverging. The left hand side of the equation would diverge from the right hand side unless $\beta_1=1,\beta_2=0$, provided that the residual $\varepsilon_t$ is required to be I(0). Proceeding with such a model would make little sense.
After differencing, you will be dealing with stationary series, and the problem will be solved.
The above presumes the variables are truly I(1). If there is some doubt about that and you are interested in impulse responses, see Ashley & Verbrugge (2009) for the idea of using VAR in levels augmented with lags of first differences.