Our ability to capture and remove carbon dioxide (CO2) at gigaton scale within a decade or two depends on our ability to quickly identify new materials that are high performing, selective over other gases with low energy demand and then further develop them for large scale deployment. As a proven technology for gas separation in other industrial applications, metal-organic frameworks (MOF) come in virtually unlimited number of crystal combinations in their highly porous lattice and may offer the solution for CO2 capture from atmosphere or industrial point sources. Although MOFs can have highly complex crystal structure, which cannot be easily exploited in tabular data format in conventional ML methods or more recent Deep Learning methods, Graph Neural Networks can easily be trained on their representative crystallographic information file (CIF) content. In this work, we train GNNs to create an end-to-end workflow to screen large number of MOF crystal structures directly from the data within the crystallographic information files for their CO2 working capacity or CO2/N2 selectivity under low-pressure conditions. Our preliminary results show that a simple 2-layered Graph Convolution Networks (GCN) can easily achieve R2 score in the range of 0.87 to 0.89, easily.