Large language models have changed the landscape for code completion, but the techniques by which prompts are constructed and model responses are handled fail to fully leverage the wealth of structured semantic information available to modern language servers and from there, IDEs. We contend that AIs need IDEs, too! This paper integrates LLMs into the Hazel Assistant, which uses static retrieval and error correction techniques driven by the Hazel Language Server. The Hazel Language Server is able to identify types and typing contexts for holes in program sketches, with total error correction ensuring that a semantically meaningful program sketch is always available. This allows for semantic contextualization with information that is not lexically local to the cursor, but is semantically local to the programmer’s intent as expressed through the type system. Contextualized model completions are then iteratively refined via dialog with a language server, providing static error correction. To validate these methods, we present MVUBench, a dataset of MVU applications written in Hazel and TypeScript, particular suited to contrast methods of static retrieval and contextualization. Further, we duplicate our methods and test suite in TypeScript in order to validate the applicability of these methods to higher-resource mainstream languages.