Models & privacy
Kiki is powered by an AI model, and by default that model runs on your own hardware. A fresh device does no outbound inference, costs nothing to run, and keeps your prompts and data local. You don't set any of this up — it's how Kiki ships.
On your device, by default
Out of the box, Kiki uses a model that runs locally and is sized for your machine. That local model is the only one used unless you decide otherwise, so:
- Your prompts, files, and screen never leave the device for inference.
- It works fully offline.
- There's nothing to pay for and nothing to configure.
Bigger machines can run bigger local models; smaller ones run a compact default and lean on the cloud only when you ask them to.
Opting into the cloud
Some tasks want a larger model than your device can run. When you want that, you turn on cloud inference — it's a switch in Settings, off until you flip it. With it on, Kiki can route heavier requests to a cloud model (Anthropic, OpenAI, or Kiki's own hosted models) while everything else stays local.
Even with the cloud allowed, Kiki keeps sensible guardrails on by default — voice stays on-device, and it falls back to the local model on low battery — so "allow the cloud" never means "send everything."
What it costs
On-device inference is always free; it's your hardware. Cloud inference, when you opt into it, is metered as part of your plan. Because Kiki is local-first, a device — or a whole fleet — can run with little or no AI cost. Cloud spend is something you choose, not something that creeps up on you.
Privacy, end to end
This is the same principle that governs the rest of Kiki: your data and what Kiki remembers about you stay on your device unless you explicitly sync or send them. Connecting to the cloud for remote control doesn't change that — inference stays local until you allow it.
Running headless devices?
On servers and edge devices you operate over SSH, the same defaults apply. Operators can change a device's model and policy from the command line; across a fleet, settings are delivered centrally from the dashboard.