Provider vault
Presets for Ark Plan, Ark Platform, Agnes, Anthropic, OpenAI and OpenSpeech. Credentials are sealed with AES-256-GCM envelope encryption, shown masked, rotatable, and verifiable with an L1 connection test.
CONTROL PLANE · NOT A GATEWAY
One control plane for every model your projects run.
Model Hub is a self-hosted, single-owner control plane. Register provider credentials once, maintain one catalog across five modalities, and assign models to projects many-to-many. Each project pulls its own projection with a token — then calls the provider directly. Inference traffic never passes through the hub.
CAPABILITIES
The four things the hub actually does
Everything else — inference, streaming, billing — stays where it belongs: inside your projects.
Five-modality catalog
Text, image, video, audio and embedding models in one catalog: multi-dimensional pricing, endpoint and dialect adapters, LiteLLM-compatible naming. Preset entries are opt-in under a strict overlay.
Assignments & budgets
Models map to projects many-to-many, with primary/fallback roles, per-project aliases and daily budgets. Over-budget calls route to a fallback, trip a circuit breaker, or stop at the kill switch.
Distribution & resilience
Projects pull GET /pub/v1/models with a Bearer token; revisions ride ETag and 304. A 0600 local cache keeps every project running even when the hub itself is down.
BOUNDARY
A control plane, not a gateway.
The hub decides which models a project may use and hands out the projection. It never sits in the request path.
- No inference traffic passes through the hub
- No streaming proxy, no centralized billing
- End-user BYOK keys stay inside each project — the hub never sees them
- Usage is reported back for accounting and budgets only
WORKFLOW
From credential to running project
Six steps, each auditable — the same loop the owner's projects already run on.
- 01
Register providers
Store credentials encrypted; verify with an L1 connection test.
- 02
Curate the catalog
Enable models across five modalities, with pricing, endpoints and dialect adapters.
- 03
Register projects
Each project gets a hashed token — the plaintext is shown exactly once.
- 04
Assign with guardrails
Many-to-many, roles, aliases, daily budgets. One batch save bumps the catalog revision.
- 05
Pull and run
model-kit fetches the projection (ETag/304), caches it 0600, calls providers directly, and reports usage back.
- 06
Watch the board
Costs by project × source × day × model, budget watermarks, and L1–L4 health checks.
THE CONSOLE
An ops surface, not a mockup
Seven screens cover the lifecycle — providers, catalog, projects, assignments, costs and knowledge — drawn with the same console-ui tokens the hub itself ships.
model-hub · admin console
DashboardProvidersModelsProjectsAssignmentsCostsKnowledge
GET /pub/v1/usage/summary — a real response shape
{
"project": "zeus",
"day": "2026-10-31",
"models": [
{
"alias": "noul",
"role": "primary",
"model": "doubao-seed-1-6-250615",
"spent_micros": 420000,
"daily_budget_micros": 1000000,
"percent": 42
}
],
"unassigned_spent_micros": 0
}WHO IT'S FOR
Built for one owner running many projects
Not a multi-tenant SaaS. A single-operator control plane for a fleet of projects that share providers and budgets.
Several AI projects at once
zeus, loom, atlas, agent-world and job-agent already read their model configuration from a single hub.
Swap models without redeploys
Change the assignment, bump the revision — projects pick the new projection up on their next refresh cycle.
Guardrails, not invoices
Daily budgets per assignment, fallback routing and a kill switch. BYOK traffic is metered, never billed.
Self-hosted by default
Single owner, docker compose up, bound to 127.0.0.1 — designed for internal or Tailscale exposure behind Caddy HTTPS.
Run your own model control plane
One compose file brings up the hub, the admin console, and a backup sidecar with a 7-day retention.
docker compose up --build # binds 127.0.0.1:8791