Brain-machine interfaces (BMI) have vastly used a single scale of neural activity, e.g., spikes or electrocorticography (ECoG), as their control signal. New technology allows for simultaneous recording of multiple scales of neural activity, from spikes to local field potentials (LFP) and ECoG. These advances introduce the new challenge of modeling and decoding multiple scales of neural activity jointly. Such multi-scale decoding is challenging for two reasons. First, spikes are discrete-valued and ECoG/LFP are continuous-valued, resulting in fundamental differences in statistical characteristics. Second, the time-scales of these signals are different, with spikes having a millisecond time-scale and ECoG/LFP having much slower time-scales on the order of tens of milliseconds. Here we develop a new multiscale modeling and decoding framework that addresses these challenges. Our multiscale decoder extracts information from ECoG/LFP in addition to spikes, while operating at the fast time-scale of the spikes. The multiscale decoder specializes to a Kalman filter (KF) or to a point process filter (PPF) when no spikes or ECoG/LFP are available, respectively. Using closed-loop BMI simulations, we show that compared to PPF decoding of spikes alone or KF decoding of LFP/ECoG alone, the multiscale decoder significantly improves the accuracy and error performance of BMI control and runs at the fast millisecond time-scale of the spikes. This new multiscale modeling and decoding framework has the potential to improve BMI control using simultaneous multiscale neural activity.