A coding agent sees more of you than a browser does. It reads your project, your credentials, your unfinished ideas. Most products answer this with a policy. Slate answers it with an architecture, because a policy is a promise and an architecture is a fact.
The model runs on the NPU. The harness runs on the CPU. The filesystem is local storage. Put the phone in aeroplane mode and every single agent function still works, which is the only honest test of the claim.
Slate reports that a run happened, so we know which tools to improve. That report has six fields and there is no field for your text:
intent build:tip
planner on-device
steps 6
tools fs.write, build, preview, git.commit
ms 41
ok true
The endpoint that receives this has no column for a goal, a filename or file contents. Sending them would fail validation rather than be quietly stored. It can be turned off in Settings, and turning it off changes nothing about how the phone works.
Once your work is on someone else's machine, you are trusting a retention schedule, a breach response and a future change of ownership. None of those are things you control. The only version of this you can actually verify is the one where the data never moves.
A resident 14 B model is not a frontier model. It is worse at long reasoning chains and it will lose to a datacentre on hard problems. Slate is a bet that a good enough model with your whole project in context, instantly, in your hand, beats a better model you have to upload your life to. For plenty of work that bet is wrong, and it is worth knowing that before you buy one.