My preferred approach for doing this is to have a hand-rolled SDK generator that reads the request, response and error models out of the microservice project and emits the same in each language targeted by the SDK, along with a minimal stub that calls the API.
You then spend 15 minutes at most, customizing the stub if needed, if you need custom behaviours like streaming.