Setup

From a clean clone, one command starts an interactive wizard:

bun run setup

The wizard:

1. Writes deploy/.env (mode 0600) with your publish host and port.
2. Optionally installs dependencies, builds images, pulls OpenViking, and starts Compose.
3. Prints the URL of your instance.

It does not ask for an AI provider — chat models are connected later, per user, inside the desktop.

What it asks

Both can be an IPv4 address or hostname. The Compose file maps ${PUBLISH_HOST}:${PUBLISH_PORT} to the control plane's internal port 8787.

First login

Open the printed URL (default http://100.96.0.3:8787) and sign in with:

admin@localhost / changeme

The first login forces new credentials, stored hashed in sqlite.

Connect a chat model

Chat models are per user, connected inside the desktop:

1. Open a thread (the desktop starts on demand).
2. On the Providers page — or via a terminal on the thread screen — run opencode auth login inside the user container.
3. Pick the model per thread from the model selector.

Configure OpenViking memory

Set the OpenViking embedding and VLM providers on the Config page. The full walkthrough — including the Cloudflare Workers AI option — is in OpenViking setup.

Manual Compose

Everything the wizard does, by hand:

bun install
bun run images                                        # build control + desktop images
docker pull ghcr.io/volcengine/openviking:latest
docker compose --env-file deploy/.env -f deploy/compose.yml up --build -d