Current methods to reconstruct muscle contributions to joint torque usually combine electromyograms (EMGs) with cadaver-based estimates of biomechanics, but both are imperfect representations of reality. Here, we describe a new method that enables online force reconstruction in which we optimize a "virtual" representation of muscle biomechanics. We first obtain tuning curves for the five major wrist muscles from the mean rectified EMG during the hold phase of an isometric aiming task when a cursor is driven by actual force recordings. We then apply a custom, gradient-descent algorithm to determine the set of "virtual pulling vectors" that best reach the target forces when combined with the observed muscle activity. When these pulling vectors are multiplied by the rectified and low-pass-filtered (1.3 Hz) EMG of the five muscles online, the reconstructed force provides a close spatiotemporal match to the true force exerted at the wrist. In three separate experiments, we demonstrate that the technique works equally well for surface and fine-wire recordings and is sensitive to biomechanical changes elicited by a modification of the forearm posture. In all conditions tested, muscle tuning curves obtained when the task was performed with feedback of reconstructed force were similar to those obtained when the task was performed with real force feedback. This online force reconstruction technique provides new avenues to study the relationship between neural control and limb biomechanics since the "virtual biomechanics" can be systematically altered at will.