It is common in applied settings for combinatorial optimization problems to be solved repeatedly with small variations. Solvers traditionally treat every new problem as novel, with no memory of solving past instances. This state of affairs had led to a surge of interest, both within academia and industry, in using machine learning methods to take advantage of the statistical relationships between recurrent problems and improve solving efficiency. However, MILP solvers are intricate pieces of software, and so far applications of machine learning methods to these solvers have been large engineering efforts that are impractical for industry. In this talk we present Ecole, a new C++ and python library that aims to simplify the development of machine learning methods in exact solvers. It offers a clean API, modeled on the well-known OpenAI Gym machine learning library, to interface with SCIP, a state-of-the-art open source solver. Ecole brings a natural separation between solver code and machine learning code, is well-tested and offers sensible defaults. This in turn makes for faster development, improved reproducibility, less bugs and reduced resource requirements. We illustrate usage of the library with ongoing projects.