Docs

Deployments

Request a GPU here or on GCP, then call it as custom:<id> through the same OpenAI-compatible API.

Dedicated deployments are not serverless. Dashboard → Deployments → Request GPU writes a row and starts a local Ollama/Metal runtime or a GCP Cloud Run GPU job, depending on the target you pick.

Targets

Target	What happens
Local	Detects GPU/Metal on this machine (`/api/gpu/status`) and runs the selected open model via the local runner
GCP	Uses the deployer in the dashboard to stand up the revision you configured

Statuses you will see (pending, building, running, stopped, failed) come from that row — Pause / Play / Delete call PATCH / DELETE /api/deployments/:id.

Calling it

Once running, the catalog exposes:

custom:<deployment-uuid>
the runtime tag (for example an Ollama name) when set

bash
curl "$GATEWAY/v1/chat/completions" \
  -H "Authorization: Bearer $OPENDOOR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "custom:YOUR_DEPLOYMENT_ID",
    "messages": [{"role":"user","content":"hello"}]
  }'

LoRA attach and routers are documented under Dedicated deployments.

vs serverless

Serverless / vendor ids in the catalog do not need a deployment. warming and dedicated rows do. Ingest with bun run ingest:open-models, then request a GPU only for models that say they need one.

Agents

Billing