Given a graph learning task, such as link prediction, on a new graph dataset, how can we automatically select the best method as well as its hyperparameters (collectively called a model)? Model selection for graph learning has been largely ad hoc. A typical approach has been to apply popular methods to new datasets, but this is often suboptimal. On the other hand, systematically comparing models on the new graph quickly becomes too costly, or even impractical. In this work, we develop the first meta-learning approach for automatic graph machine learning, called AutoGML, which capitalizes on the prior performances of a large body of existing methods on benchmark graph datasets, and carries over this prior experience to automatically select an effective model to use for the new graph, without any model training or evaluations. To capture the similarity across graphs from different domains, we introduce specialized meta-graph features that quantify the structural characteristics of a graph. Then we design a meta-graph that represents the relations among models and graphs, and develop a graph meta-learner operating on the meta-graph, which estimates the relevance of each model to different graphs. Through extensive experiments, we show that using AutoGML to select a method for the new graph significantly outperforms consistently applying popular methods as well as several existing meta-learners, while being extremely fast at test time.