Why local is a structural guarantee, not a policy
Cloud vendors publish privacy and data-retention policies, and many are good. But a policy is a promise you have to trust, and it doesn't change the fact that your code physically travelled to someone else's servers. For sensitive work, the safest property is the one you don't have to trust: the data never left.
Running the model locally gives you exactly that. oi talks to a runtime on localhost, the weights are on your disk, and inference happens on your CPU or GPU. There's no remote endpoint in the path, so there's nothing to leak, log, or subpoena — privacy by architecture rather than by agreement.
Air-gapped and offline by default
Because nothing needs the network at inference time, a local assistant works on an air-gapped machine, on a plane, or on a locked-down corporate network where outbound traffic is blocked. Once the runtime and model are installed, you can pull the plug and keep working — completion, chat, refactors, all still there.
That's a capability cloud tools simply can't match: a cloud model with no connection is a model that doesn't run. For teams who can't or won't open an outbound path for their editor, local isn't a downgrade, it's the only thing that works.
What you trade, honestly
Local-first isn't free of tradeoffs. You take on the hardware, and open models won't match a frontier cloud model on the very hardest problems. The case here is narrower and stronger: when the requirement is that code cannot leave the machine, capability is secondary — a good-enough model you're allowed to use beats a great one you aren't. oi is built for that constraint, free and local by default.
frequently asked
Does any of my code reach the internet?
No. oi connects to a local runtime on localhost, so prompts and source stay on your machine. After the one-time model download, inference makes no network calls.
Will it work on an air-gapped machine?
Yes. Once the runtime and model are installed, nothing needs the network at inference time, so it runs fully air-gapped or offline.
Is this enough for NDA or regulated codebases?
Keeping code on-device removes the main exposure of cloud AI — that your source is sent to a third party. Always confirm against your own compliance requirements, but local inference is the architecture built for exactly this constraint.
Do I give up capability for privacy?
Somewhat — open local models trail frontier cloud models on the hardest tasks. But for the everyday majority they're good enough, and when code can't leave the machine, a usable local model is the only real option.
Last updated June 19, 2026