Skip to content

Networking

Two separate overlay networks do two separate jobs here — easy to conflate, worth keeping distinct:

  • Tailscale connects the hosts to each other.
  • Cloudflare (Tunnel, Access, DNS, Pages) connects the outside world to Lana — nothing else. Cheryl is outside this model: its game ports, Mattermost (HTTPS) and Mumble are exposed directly on its public IP, filtered by Proxmox’s firewall on figgis.

The servers share one tailnet:

Tailscale name Host Tailscale IP
lana Lana, VM on figgis 100.112.192.98
cheryl Cheryl, VM on figgis 100.107.153.62
figgis figgis, the Proxmox host 100.101.130.59
krieger Krieger 100.90.162.24
pam Pam 100.109.83.48
woodhouse Home Assistant box 100.107.99.26
lana-old, cheryl-old the old VPSes, being retired (nothing live) 100.115.215.23, 100.67.110.77

Lana’s wake-proxy containers reach Krieger’s Ollama/ComfyUI over the tailnet, with no home-network port forwarding (see Krieger). Lana is the exit node (since 2026-11-18, tailscale_exit_node in the playbook): traffic through it leaves the internet from figgis’s Hillsboro IP 147.135.38.231, and clients reach it directly, not through a relay. Old Lana no longer offers it.

There are a couple of other tailnet members (a phone, this Mac) not relevant to the server-side stack. Their Exit Node setting should normally be None: with Lana selected (or the “auto” recommended exit node, which can only pick it), all their web traffic leaves from a datacenter IP, which sites often block or challenge. On 2026-11-02 the Mac had both set. The tailnet pushes no DNS resolvers and no subnet routes, so with no exit node only 100.x traffic uses Tailscale.

The tailnet’s ACLs are port-specific, not “everything can reach everything”. The policy lives in docker-infra/tailscale/policy.hujson (grants syntax, commented) and is applied by a Gitea workflow on push; editing it in the admin console is disabled (since 2026-11-17). Lana is tag:lana (lana-old also carries tag:exit-node), Cheryl tag:cheryl (the VM and cheryl-old, which runs nothing live since 2026-11-17), figgis tag:figgis, Pam tag:pam, the Home Assistant box tag:woodhouse. The paths that aren’t obvious:

From To Port For
Lana cheryl TCP 8080 The Woodhouse bot calling the AMP API (published on the tailnet by tailscale serve)
Lana tag:cheryl TCP 22 DOCKER_HOST=ssh://cheryl: deploying Mattermost and Mumble (the key only allows docker system dial-stdio)
Lana tag:cheryl, Pam TCP 7007 Dozzle reading its agents
tag:cheryl Pam TCP 3900 AMP uploading backups to Garage’s S3 API
tag:external-vpn Pam TCP 8085, 8989, 7878, 9696 The arr web UIs directly
members, tag:external-vpn tag:cheryl TCP 8080 The AMP panel
members, tag:external-vpn tag:figgis TCP 22, 8006, 3128 figgis: SSH, Proxmox web UI, consoles
members, tag:external-vpn svc:* TCP 80 Short names for the web UIs (Tailscale Services, below)

Web UIs have names of their own: http://nzb/, sonarr/, radarr/, prowlarr/, grafana/, jellyfin/, homeassistant/, logs/, amp/, nas/ (or <name>.taila2470a.ts.net), and https://proxmox.taila2470a.ts.net (HTTPS with Tailscale’s certificate, which only covers the full name; http://proxmox/ serves a small page that redirects there). Each is a Tailscale Service with its own tailnet address, so all use port 80.

Service Hosted by How
svc:homeassistant Home Assistant box Tailscale add-on services option, TCP to 127.0.0.1:8123
svc:jellyfin, svc:nzb, svc:sonarr, svc:radarr, svc:prowlarr, svc:grafana Home Assistant box the same option, TCP to the “Pam relay” add-on, which forwards to Pam over the LAN
svc:logs Lana tailscale serve --service=svc:logs --http=80 127.0.0.1:8180 (Dozzle)
svc:search Lana tailscale serve --bg --service=svc:search --http=80 127.0.0.1:8190 (SearXNG); auto-approved by autoApprovers in tailscale/policy.hujson
svc:amp Cheryl --http=80 127.0.0.1:8080 (the AMP panel), from the playbook’s tailscale_services
svc:proxmox figgis --https=443 https+insecure://127.0.0.1:8006, plus port 80 serving a redirect page; both from tailscale_services (the service needs tcp:80 and tcp:443 endpoints)
svc:nas Home Assistant box the add-on option, TCP to the relay’s port 15000, which forwards to DSM on Pam:5000
  • Defined in the admin console (Services), each with endpoint tcp:80. The endpoint must be the port the host serves on, not the app’s port: with tcp:8085 clients got the DNS name but no route, and connections timed out. A host that started advertising a service before it existed has to re-advertise (tailscale serve clear, then serve again).
  • Pam can’t host a Service: its Tailscale runs in userspace mode, and the control plane never lists it as a host. Everything on Pam is therefore hosted by the Home Assistant box, which sits on the LAN beside it, so the traffic stays at home. Until 2026-11-17 old Cheryl hosted the arr names and svc:grafana (through a socat relay to Pam:3000).
  • The Home Assistant add-on clears any service host not in its services option when it starts, takes only 127.0.0.1 targets, and is set as plain TCP because Home Assistant rejects proxied X-Forwarded-For. Details in homeassistant/README.md in docker-infra. autoApprovers.services approves tag:woodhouse for these, tag:lana for svc:logs.
  • Clients need a recent Tailscale too: Sterling on 1.96.5 got no route; 1.104.6 works.

Tagged devices aren’t members: the Mac (sterling) carries tag:external-vpn, so an autogroup:member rule doesn’t cover it and it needs a rule of its own. Pam’s Tailscale runs in userspace mode, which forwards tailnet connections to its local ports; plain TCP services work through it.

Lana cannot reach Pam on 3900. That only matters if Lana is ever backed up to Garage too; it needs an ACL rule first.

Nothing on Lana has a port directly exposed to the internet. Public traffic reaches Docker-hosted apps through one path only:

Browser → Cloudflare (DNS + Access) → Cloudflare Tunnel → cloudflared on Lana → app container

DNS. The spelvin.dev zone lives in Cloudflare. Every app hostname (git.spelvin.dev, roadmap.spelvin.dev, etc.) is a proxied CNAME to <tunnel-id>.cfargotunnel.com — never an A record pointing at Lana’s real IP.

Tunnel. One named Cloudflare Tunnel (confusingly named flowviz in the dashboard, after the first app tunneled, since removed; it carries every app) runs as a cloudflared container on Lana, outbound-only. Its ingress config is a flat list of hostname → internal service:port rules, e.g. git.spelvin.dev → http://gitea:3000, terminated by a catch-all 404. This config lives in Cloudflare’s dashboard/API, not as a file in docker-infra.

Access. Sits in front of the Tunnel (and, separately, in front of Cloudflare Pages sites when needed) as the actual login gate. Two patterns are in use:

  • self_hosted Access Application — gates a hostname outright; nothing reaches the app without passing Access’s login page first. Used for apps with no native SSO of their own (none left since FlowViz and Windmill were removed).
  • saas Access Application — Access acts as an OIDC identity provider that the app itself talks to, so the app has its own login page but delegates identity to Access rather than maintaining its own user database. Used for apps with native OIDC support (Gitea, LiteLLM, Open WebUI, Leantime).

Access policies used to be inline, per-app, single-email-match rules — that model is retired. As of today, every app is gated by one of two reusable Access policies instead:

  • “Only Me” — for personal-admin-only apps (LiteLLM and the rest of the single-user apps).
  • A household policy — for apps shared with other household members.

Both match via an OIDC Claims rule against Authentik group membership (a groups claim) rather than by email address, so adding or removing someone is a group-membership change in Authentik, not an edit to every app’s Access policy. Whoever is authenticating can do so through either of two Identity Providers registered in Access, shown side by side on every login screen since no app restricts allowed_idps:

  • Cloudflare — Access’s built-in email one-time-PIN provider. The original/default option.
  • Authentik (Passkey) — a generic OIDC provider pointing at Authentik, self-hosted specifically to add passkey (WebAuthn) login. See authentik/README.md in docker-infra for the non-obvious provider settings (signing key, scope mappings, grant types) that had to be right before Cloudflare would accept it — the failure mode for all three was an identical, unhelpful “Failed to fetch user/group information from the identity provider.”

Two hostnames deliberately skip step 4 and have no Access Application at all, reachable straight through the Tunnel with their own gate:

  • mcp.spelvin.dev — knowledge-mcp, which runs its own MCP OAuth and sends the login to Authentik (Anthropic’s servers can’t pass an Access login page).
  • steam.spelvin.dev — Woodhouse’s Sign in with Steam page and its game links (/app/<id>, which open the Steam client). Friends without an Authentik account have to reach it; Steam’s own login is the gate, and a one-time link from /steam link in Discord ties the sign-in to a Discord user. The Tunnel route targets http://steam-login:8080, a network alias, because the short name woodhouse resolves to the Home Assistant box on the tailnet.

(ntfy.spelvin.dev used to be a third; ntfy was retired in October 2026.) See Services.

Authentik itself is not behind Access — it can’t be, since Access redirects unauthenticated users to it to log in. Its own login page (passkey-only) is the actual boundary there.

Pages. spelvin.dev and nothingwilts.com bypass all of the above entirely — they’re static Astro sites deployed straight to Cloudflare Pages by a Gitea Actions workflow (wrangler pages deploy) in each site’s own repo, with a custom domain attached via the same API token. No container, no Tunnel entry, no origin on Lana at all.

The recipe, in order:

  1. Deploy the container on Lana (docker-infra/<app>/docker-compose.yml), on the shared web Docker network so cloudflared can reach it by container name.
  2. Add a DNS CNAME: <hostname>.spelvin.dev → <tunnel-id>.cfargotunnel.com, proxied.
  3. Add an ingress rule to the Tunnel config: <hostname> → http://<container>:<port>.
  4. Create a self_hosted Access Application for the hostname, and attach the appropriate reusable policy (“Only Me” or the household policy — or a saas Application instead, if the app supports native OIDC and you’d rather avoid a double login).

Static sites skip steps 2–4 in favor of the Pages custom-domain flow, plus an Access Application on the Pages hostname if it needs gating too.