it was invoked manually at build time, and then at CI/CD time (there was no runtime type checking). we did output types to pyi files (the python equivalent of C header files), but mainly we used the inferred types to check the code for type correctness - one of the unique selling points of pytype, and why google was developing its own typechecker in the first place, was that the other available python type checkers would check annotated code for consistency, but would not do anything about type errors in unannotated code. if pytype did detect type errors, the code failed CI and had to be fixed before it was checked in.

we did also have a separate tool that could merge the inferred types back into the source code as annotations, but it never got much uptake.