Brain machine interfaces (BMIs) enable us to control external devices using our brain signals. Using a grid-shaped flickering line-array and a shrink-rLDA classifier, top-down information could recently be decoded in a steady-state visual evoked potential (SSVEP)-based BMI paradigm. The present study tested its feasibility in online implementation. We found that within reasonable computing time (0.114 s on average) its online system was successfully accomplished with a decoding accuracy of 53.7% on average. The accuracy was 3.2 times significantly higher than the accuracy by random-shuffled data (16.7%). Therefore, using the grid-shaped SSVEP-based BMI, one's multiclass (at least 6 classes) intention can be online decoded and subsequently control external devices.