This document presents Specformer, a novel Transformer-based spectral graph neural network. Specformer uses a Transformer to encode eigenvalues of the graph Laplacian into representations that capture magnitude and relative differences. It then applies self-attention to learn relationships between eigenvalues. Specformer decodes the eigenvalue representations using learnable bases to perform graph convolutions. Experiments show Specformer outperforms other GNNs on node and graph classification tasks and learns meaningful patterns in the graph spectrum.