In this work, we propose a new descriptor that is called Gradient Local Binary Patterns (GLBP) for automatic keyword spotting in handwritten documents. GLBP is a gradient feature that improves the Histogram of Oriented Gradients (HOG) by calculating the gradient information at transitions of the Local Binary Pattern code. For the matching step, we use the Euclidian Distance and the Cosine Similarity. To show GLBP's performance, we used a Benchmark dataset which contains 100 documents written if 4 languages, from those documents 300 query were extracted to be spotted. The results obtained highlight the effectiveness of the proposed descriptor.