Developers

The surfaces you build against

LokiAI is a control plane, a device agent protocol and a set of runtime adapters. Everything else is a consequence of those three.

01Architecture

Four moving parts

Accounts, devices, model records, deployment state and credential issue. Stateless with respect to inference.

Registers capability, holds a device-scoped credential, executes deployments and streams state.

Turns a repository reference into one pinned file with recorded format, size and URL evidence.

A narrow interface per runtime: load, warm, infer, release. Adding a runtime does not touch the control plane.

02Lifecycle

Agent authenticates over WSS with its device credential and publishes its capability profile.
Control plane resolves the requested outcome to a candidate artifact and runs compatibility checks.
On pass, a short-lived download instruction is issued to that device only.
Agent downloads directly from the CDN, verifies integrity, then loads through the matching runtime adapter.
State transitions and failure reasons are streamed back and stored with the deployment record.

03Integration

GGUF and TFLite are deployable today. Additional formats are under evaluation.

Account, device, model and deployment resources. In development — contact us for early access.

Thin clients over the same API, starting with TypeScript and Python.

Implement the agent protocol for your platform; the control plane needs no change.

Deployment state changes are streamed, with structured failure reasons rather than opaque errors.

Getting started, pairing, compatibility and troubleshooting are maintained alongside the agent.

Start with the docs, then contact us with the platform you are targeting.