asgard.wovenbytes.com
One endpoint for every model we run.
asgard is the single OpenAI-compatible gateway in front of the whole fleet: hosted cloud models, work Vertex, and the machines in the rack. One base URL, one key, any OpenAI SDK — clients never see a provider, a LAN address, or a provider key.
Sign-in is for the human pages — the fleet console and media
jobs. Programs use a client key against /v1 and never see a login.
Chat & completions
/v1/chat/completions across cloud
providers, Google AI Studio, work Vertex and models served on our own hardware. Claude Code points
at it directly — /v1/messages speaks the Anthropic API natively, thinking included.
Embeddings & rerank
/v1/embeddings and
/v1/rerank, on self-hosted bge-m3 and a self-hosted reranker, with cloud
lanes behind them for overflow.
Video & music
Long renders run as jobs: submit to
/v1/media/jobs, poll it, and the artifact comes back as a time-limited download link.
The bytes never travel through the gateway.
Images, speech, transcription
Image generation and editing, text to speech, and transcription with diarization — each translated at the edge into the same OpenAI shape as everything else.
The orchestrator
A planner that turns a request into a DAG and runs specialist agents over it. It is addressed as a model name, so any OpenAI client can drive it.
Live catalog
GET /v1/models is public and enumerated live
from every backend, so the list is what is actually served — not a baked
snapshot.
Machines authenticate with a bearer client key. People sign in with Google through Cloudflare Access; the pages above are the only paths behind that login, because everything else on this hostname is machine traffic that has no browser.