Deploy remo on a server

What you get

A working installation on one server: nginx terminates HTTPS, the sync service holds rooms and the API, Postgres stores boards and access.

One instance. Rooms, open sockets and access checks live in the process; two copies behind a load balancer will not see each other's rooms. In this version scaling is up, not out.

Before you start

Ubuntu 24.04, Node 22 LTS, pnpm 11, Postgres 16. nginx, certbot, git and ufw are required.

apt install -y nginx certbot python3-certbot-nginx git curl ufw
corepack enable && corepack prepare pnpm@11.2.2 --activate
useradd --create-home --shell /bin/bash remo
install -d -o remo -g remo /srv/remo

Steps

1. Code: sudo -u remo git clone <repo> /srv/remo && cd /srv/remo && sudo -u remo pnpm install --frozen-lockfile.
2. Database: POSTGRES_PASSWORD='<strong>' POSTGRES_PORT=127.0.0.1:5432 pnpm db:up, then pnpm db:check. The port with the host is mandatory: without it Postgres listens on every interface.
3. Environment: create /srv/remo/.env with DATABASE_URL, NODE_ENV=production, PORT=1234, REMO_APP_ORIGIN, REMO_CORS_ORIGIN, REMO_METRICS_TOKEN, REMO_KEK, REMO_KEK_ID; chmod 600.
4. Build: sudo -u remo pnpm build. The static files land in apps/web/dist.
5. Service: copy deploy/remo-sync.service to /etc/systemd/system/, enable and start it.
6. nginx: copy deploy/nginx-remo.conf, replace the domain, enable the site, nginx -t && systemctl reload nginx, then certbot --nginx -d <domain>.
7. Documentation portal: pnpm guide:build and rsync -a --delete docs-site/ /srv/remo/docs-site/; nginx serves it at /guide/.

Check

curl -s https://<domain>/api/health          # {"ok":true,…}
curl -sI https://<domain>/                    # 200
curl -sI https://<domain>/guide/              # 200, documentation

In the browser: sign-up, a board, a second tab in a private window, dragging a sticky note. In the Network panel the socket must show 101 Switching Protocols.

Limitations

• One process: a restart breaks sockets, clients reconnect on their own.
• One agent task runs at a time: a slow model delays the agent queue, not the canvas.

Where to go next

Installation security — what to close before the first user.