The hub
One device or server per home or office is the hub. Clients and apps only ever know the hub; the hub knows the nodes. The hub:
- Authenticates every request with a virtual key (
sk-aigw-…, LiteLLM style). Keys carry model-glob scopes (*,clip-*), a per-minute rate limit and a daily quota. Out-of-scope models answer 404 as if they did not exist. - Audits every request: time, key name, IP, method, path (no query), model, node, status, latency, bytes in/out, request id. Never keys or bodies.
- Routes each job to a node (see Routing).
- Speaks two APIs: the native API (raw image bytes in) and the
OpenAI-compatible API under
/openai/v1. - Keeps the registry of benchmark results reported by nodes.
The server edition runs no models itself. The phone can also act as a LAN-only hub with its own models (the original mode, with pinned self-signed certificates), but the recommended setup is a server hub with phones and TVs as nodes.
Key kinds
Section titled “Key kinds”| Kind | Can |
|---|---|
admin (AIGW_ADMIN_KEY, not stored) |
/v1/admin/* and the dashboard, no rate limit |
client |
the native and OpenAI APIs, within its scopes |
node |
only /v1/nodes/connect, /v1/nodes/register and POST /v1/registry/results |
Only HMAC-SHA256(pepper, secret) is stored, so a copied database cannot be used to test guessed keys.