OpenViking setup

Each desktop's OpenViking instance (the memory server on port 1933) needs an embedding model and a VLM. Every user sets their own on the Config page; an admin sets a fallback under Admin → OpenViking models (default) for users who have not saved their own.

Prerequisites (Cloudflare Workers AI example)

• A Cloudflare account with Workers AI enabled (Workers Paid for the models below)
• An API token with Workers AI edit permission (dash.cloudflare.com → My Profile → API Tokens)
• Your Cloudflare account ID (dashboard overview, right sidebar)

Set the models

1. Log in to the control plane.
2. Open Config (per user, saves to your own desktop) or Admin → OpenViking models (default) (account-wide fallback).
3. Fill the form:

4. Save OpenViking models — a blank API key keeps the already-saved key.

The Base URL is the Workers AI OpenAI-compatible endpoint; model IDs keep the @cf/ prefix.

Apply

Sleep the desktop, then start it again — desktop start recreates the per-user OpenViking container with the saved config passed via the OPENVIKING_CONF_CONTENT env var (apps/control/src/docker.ts, vikingConfig). There is no config file to manage on the host.

Form limitations

The Admin form sends no extra fields, so it cannot set:

• cf-aig-gateway-id — the host ~/.openviking/ov.conf can route embeddings through an AI Gateway (named embeddings) for caching and logs. Without it, requests go straight to Workers AI and still work.
• reasoning_effort: low for the VLM — GLM-5.3-Flash answers without it.

Verify

curl http://127.0.0.1:8787/api/health          # {"ok":true}

• Start a desktop, then from inside the OpenCode container: ov find "anything" — results mean embeddings work.
• Ask the agent to take a screenshot — a described image means the VLM works.