Don't forget making the step of making the initial determination of whether something is a command or not happen locally. We don't need any more always-on surveillance devices. Even a new home-based intelligent assistant marketing itself as a privacy-minded alternative doesn't include such a basic safeguard.
Definitely an open, web-based architecture is a good thing. Most popular assistants will likely need to make their own determination of which service, which API, to talk to to get the response you want; very few people will want to keep track of many options and what to call them and when one gives a better result.
So it's worth looking at what this mediation layer would look like, what kind of machine learning could be used, and how to make it as free and hackable (by the user, not others) as possible.