Every cloud coding assistant sends your source somewhere to be processed. For most software that is an acceptable trade. For a studio sitting on an unannounced title under NDA, it is the entire conversation.
What a studio is actually worried about
Retention. How long is the code kept, and in what form? The answer is in a policy that can change.
Training. Is your project used to improve the model? Most providers now say no for business tiers. "Most" and "now" are both doing work in that sentence.
Contractual exposure. If you are working under a publisher's NDA, you may have already agreed not to transmit the code to third parties. That makes the question legal rather than preferential.
Breach. Any system holding source is a target. Yours would be somewhere inside it.
What local actually buys you
A model running on a machine inside your own network removes all four at once — not by having better policies, but by removing the transmission. There is nothing to retain, nothing to train on, nothing to disclose and nothing to breach.
It also changes the cost shape. Cloud assistants bill per seat, every month, indefinitely. A local setup is hardware you own, and the bill does not grow with the team.
What you give up
Honesty requires this part: local models are generally smaller than the largest hosted ones, you own the hardware and its maintenance, and updates are a thing you do rather than something that happens. For some work the hosted model is simply better, and pretending otherwise helps nobody.
This is what we do day to day — the AI Game Generator at Factorial Studio.
