Platform

One control plane. Many device agents. No cables.

The platform separates decision-making from execution. The control plane decides what should run where; the agent on the device does the running.

01Subsystems

Eight parts, one deployment path

Each subsystem has a narrow job. A deployment fails loudly at the first one that cannot proceed.

Holds accounts, devices, model records and deployment state. Issues credentials, never handles inference data.

Per-platform clients that report capability, receive instructions and execute deployments locally.

Resolves a repository to one exact artifact and records format, size, runtime and URL evidence.

Reads chipset, memory headroom, accelerators, OS version and runtime support from the device itself.

Chooses the adapter that can actually load the artifact on that profile, or refuses.

Sequences credential issue, direct download, integrity check, load and activation.

Streams state transitions and failures per deployment, retained as an auditable record.

Resumable transfers, safe re-pairing, credential revocation and rollback to the last good artifact.

02Agent surface

Adding hardware means adding an agent, not rewriting the platform.

LokiAI Control PlaneDiscovery · Verification · Deployment · Telemetry

03Boundaries

It does not proxy model weights — devices download directly from the source CDN.
It does not receive inference inputs — prompts, frames and audio stay on the device.
It does not return raw provider secrets to the browser.
It does not silently substitute a different artifact than the one recorded.
It does not assume hardware capability that the agent has not reported.

The developer overview covers the agent protocol, runtime adapters and deployment events.