What changes about data flow#
With a hosted endpoint, the prompt context your client sends reaches the vendor. That is the integration working as designed, but it is the fact your data-protection review needs stated plainly.
With a self-hosted package, traffic stays inside your network and the vendor sees only whatever the server itself calls out to — which may still be their API.
Self-hosting a server that proxies a vendor's cloud API changes where the code runs, not where the data ultimately goes. Establish which shape you are actually dealing with.
What changes about credentials#
Hosted servers typically use OAuth, so credentials are delegated, scoped and revocable, and your IdP can participate.
Self-hosted servers typically need an API key in their environment. That key lives in your secret store and is your responsibility to rotate — more control, and more to operate.
Check which inputs a package declares and whether credential-shaped ones are flagged as secrets. An unflagged token input can end up in a log.
What changes about failure#
A hosted endpoint fails as an external dependency: you observe it, you do not fix it, and you have no visibility into why.
A self-hosted server fails as your workload: you fix it, and you own the pager.
Neither is better in the abstract. The question is which failure mode your organisation is actually equipped to absorb.