A closed-loop brain-machine interface for elbow assistance is proposed in this work. The system decodes flexion and extension movements from noninvasive electroencephalographic signals through Kalman filter and uses them to activate a virtual device. A two-degree-of-freedom control scheme drives the model through a decoded path by generating a set of estimated inputs using differential flatness. These inputs are compensated by a feedback loop when decoding errors or external forces actuate the model. The results provide us an insight into the control architecture performance, which is dependent on the decoding precision. These decoding capabilities can be manipulated through a set of parameter configurations, enhancing the path tracking, or decreasing the decoding error according to their values.