This paper proposes a variable-Horizon based model predictive control (MPC) to compensate for a system with timevarying delay. The proposed method makes use of a prediction model to estimate a controlled object's future state. Suppose time delay varies, the local-side prediction assumes the worstcase delay Most packets however arrive within the worst-case maximum delay. Waiting for the timing until the maximum time runs out is inefficient. Instead, this paper proposes using the early-arriving packets to produce the remote-side reference trajectory for MPC. This paper's contribution is a proposal for a framework for MPC to deal with varying horizons. The method's performance was evaluated by numerical simulations and remotecontrol experiments with a human operated mobile robot.