Leaving Tailscale for plain WireGuard: one hub, one DNS name, help only on request
In plain English. Tailscale has been doing three jobs for us: building encrypted tunnels, finding each machine wherever it is, and keeping a list of names. Plain WireGuard only does the first job. So moving off Tailscale means picking a replacement for the other two.
- Finding machines: make one box the meeting point (our UniFi gateway) and have every other machine call it. Only that one box needs a stable name. Its home internet address changes from time to time, so dynamic DNS keeps the name pointed at it.
- People we help: their Macs call our gateway, never the other way round, and only when they want help. Their tunnel is off the rest of the time, so nothing on their Mac is waiting for us. Their changing address does not matter either: WireGuard learns it from their packets.
- The catch on Macs: the macOS WireGuard app does not look the hub's name up again while a tunnel stays up. For tunnels switched on fresh each time, that matters much less.
- Remote help: the built-in macOS Screen Sharing, carried inside the tunnel, switched on by the person asking for help. No third-party remote-desktop app.
- Outside collaborators get their own tunnels. Each tunnel opens exactly one door: the GPU machines for one group, the Forge server for another. The gateway firewall enforces that, not the person's own config file.
- What we give up: direct machine-to-machine paths, a relay for networks that block UDP, automatic names, and access rules managed in one place.
Bottom line: it is doable with gear we already own. Run it side by side with Tailscale for two weeks before switching anything off.
Figure 1. Target shape. The dashed box is the only name that has to follow a changing IP address.
What we have today
Checked from the M4 Max on October 31, 2026. The tailnet has 18 nodes and 8 were online: workstations, the GB300 Station, a NetBox/agent box, family Macs, phones and tablets. The lab machines and the family Mac Minis all sit on the home LAN with the gateway, so they need no tunnel at all; they stay reachable on the LAN exactly as today. Tailscale only mattered when James was away from home.
Both of our sites run UniFi gateways on UniFi OS 5.1.132. The main gateway holds a real public IPv4 address. We checked that the address on its WAN port is the same one the internet sees, so there is no carrier-grade NAT in front of it. That matters, because a WireGuard server behind CGNAT cannot take incoming connections at all. The address is assigned by DHCP, so it can change.
The dynamic DNS problem
A WireGuard peer can point at a hostname, but WireGuard looks that name up once and then never again (ArchWiki). DDNS only helps if two things happen. The name has to stay current. And clients have to look it up again when the address changes.
Keeping the name current
- Our blog domain's DNS can't be updated by a script. It is hosted at GoDaddy. Since May 2024 GoDaddy only gives DNS API access to accounts with 10 or more domains or a paid Discount Domain Club plan (summary with the vendor notice). We don't qualify.
- UniFi has a DDNS client built in. It is inadyn. Officially supported providers are afraid, cloudflare, dnspark, dslreports, dyndns, easydns, namecheap, noip, sitelutions and zoneedit. It only works when the gateway has a public IP (Ubiquiti).
- Pick: buy a cheap domain just for infrastructure and host it on Cloudflare's free plan. Make a Cloudflare API token that can edit DNS in that one zone. Point the gateway's DDNS at it, using the WAN that actually carries traffic. This leaves the blog and its mail records alone. A zero-cost alternative is a free afraid.org name plus a
CNAMEfrom the existing domain. - Belt and braces: also run favonia/cloudflare-ddns in a container on a LAN server. Community posts say inadyn on UniFi can stop updating after a gateway restart. We have not reproduced that. A second updater is cheap insurance.
- Watchdog: a Hermes no-agent cron job every 10 minutes compares the name to the public IP seen from the LAN. It sends an email if they disagree for more than 15 minutes. Keep the record TTL at 60–120 seconds.
Making clients look the name up again
- Linux and Homebrew
wg-quick: runreresolve-dns.shfrom wireguard-tools every 30–60 seconds. - iPhone: the iOS app sets its peer endpoints again on every network change (per the description in the PR below), so a phone moving between Wi-Fi and cellular picks up a new hub address at the next switch.
- macOS app: on a network change it only rebinds its sockets and keeps the old endpoint. The fix is in wireguard-apple PR #63, which had not been merged as of this writing. A Mac that keeps one tunnel up for days never re-resolves. A tunnel switched on fresh resolves the name at that moment.
Remote help, only when they ask
Some of the people we look after want occasional admin help on their Macs, with their permission, and nothing more. Their Macs should not accept connections the rest of the time. Plain WireGuard fits that well, because the person being helped decides when the door exists.
- Tunnel off by default. They use the App Store WireGuard app, which puts a toggle in the menu bar. When they want help they switch it on. When we're done they switch it off. While it's off, nothing of ours can reach their Mac.
- Screen Sharing off by default too. The tunnel is not the only listener. Screen Sharing, when enabled, also answers on their own home network. The simple version is two clicks for them: the WireGuard toggle and System Settings → General → Sharing → Screen Sharing. A nicer version is one "Get help" shortcut that turns on both and a "Done" that turns both off, with an automatic off after a few hours in case they forget. We will build and test that on one of our own Macs first.
- When it's on: we connect with Screen Sharing.app or Finder → Go → Connect to Server →
vnc://their tunnel address. Screen Sharing allows one named support account, not "all users". Remote Management stays off; Apple does not allow both at once. - Consent is built in. They are at the Mac, they turned it on, and they can end the session from the menu bar.
- Their side routes only our admin addresses: the admin tunnel subnet
10.66.0.0/24plus the one home Mac we help from, as a/32. Without that/32, replies to a session started from our home LAN would try to leave through their own internet. Their internet and home LAN are untouched. - Firewall it one way. Their tunnel lands in its own zone. Our admin devices may open TCP 5900 (Screen Sharing) and 22 (SSH) to them. Their side can only reply. If their machine is ever compromised, it should not open a door into the lab.
- One tunnel per person on a Mac. The macOS app runs one tunnel at a time (lifting that is part of the same unmerged PR #63). Someone who is helped and also uses a lab server gets one peer whose zone carries both rules, instead of two tunnels.
- No DDNS at their end and no open port on their router. They always dial out.
- Keep the current remote-desktop app as break-glass until a few real help sessions have worked this way. Then retire it (below).
Outside collaborators: one door each
Two groups of outside people need in. One group needs shell access to the GB300 Station and the DGX Sparks. The other needs the Forge server. The trap is UniFi's default. In its zone-based firewall, every WireGuard client lands in the built-in VPN zone, which is Allow All to the internal LAN, the gateway and other VPN clients. A collaborator on the default setup could reach the vault, the NAS and the Macs of people we help.
| WireGuard server | Who | Firewall zone | May reach |
|---|---|---|---|
| admin · UDP 51820 · 10.66.0.0/24 | James's devices | built-in VPN | Everything, like being home |
| help · UDP 51821 · 10.66.1.0/24 | People we help, on request | custom "Help" | Nothing (reply only), plus a lab server for anyone who also works on one; reachable on 5900/22 from admin while their tunnel is on |
| gpu · UDP 51822 · 10.66.2.0/24 | GPU collaborators | custom "Guests-GPU" | SSH (TCP 22) to the Station and the eight Sparks; nothing else |
| forge · UDP 51823 · 10.66.3.0/24 | Forge collaborators | custom "Guests-Forge" | Forge on the ports they actually need; nothing else |
- Why separate servers instead of per-person rules: a separate server gives each group its own subnet and zone, and a zone rule is easy to audit. There is also a community report on UniFi Network 10.3.55 that a VPN policy keyed on one client's source IP matched every client on the same WireGuard server. We have not reproduced it on our 10.6.109, and with this design we don't depend on it.
- Each custom zone: block all to Internal, Gateway, VPN and the other custom zones, then add allow policies for the exact hosts and ports. Blocking the Gateway zone also keeps them off the UniFi admin page and the gateway's DNS. Their configs carry IPs and no
DNS=line. - Client side: each config's
AllowedIPslists only that person's target hosts as/32s, so their own internet and home LAN are untouched. A/32route wins even if their home network is also192.168.1.x. The client config is a convenience. The gateway firewall is the real boundary. - Delivery: UniFi generates each config with its private key inside. Hand it over as a one-time, expiring Vaultwarden Send or a QR code in person, never by email or chat.
- On the hosts: one Linux account per person with their own SSH key, password login off, and no sudo or docker group by default. Docker group membership is root in disguise. The GB300 also serves our live interactive model, and a guest GPU job can starve it. GPU time on the Station needs an explicit rule, not just a login.
- Offboarding is one click: delete the peer and the door is gone. WireGuard keys never expire, so a monthly Hermes reminder will list every collaborator peer with its last handshake. Stale ones get removed.
Retiring third-party remote-desktop apps
Once the tunnels work, nothing needs a vendor's cloud relay to reach our machines. Macs get built-in Screen Sharing over WireGuard. Linux machines get SSH. A Windows machine, if one turns up, gets RDP over WireGuard. For each machine with a third-party remote-desktop app: prove the replacement path, turn off the app's unattended access, wait two weeks, then uninstall. Unattended access through a vendor relay is an internet-facing way in, so removing it is a security gain as well as one less subscription.
What we give up
| Capability | Tailscale | Plain WireGuard hub |
|---|---|---|
| Path between two remote machines | Direct, via NAT traversal | Through the hub, both ways |
| Networks that block UDP (some hotels) | DERP relay over HTTPS | Tunnel fails; keep a fallback |
| Names | MagicDNS | Gateway local DNS, pushed as the tunnel's DNS server |
| Who can reach what | Central ACL policy + identity | Gateway firewall rules per tunnel IP |
| Keys | Rotated and expired by the service | Ours: generate on device, vault any config that holds a private key |
| Extras (Taildrop, Tailscale SSH) | Built in | Gone; plain SSH and AirDrop instead |
| Third-party coordination service | Required | None |
Middle path. If the main goal is getting off Tailscale's cloud rather than off any coordinator, Headscale self-hosts the control plane. It keeps the Tailscale clients, NAT traversal and MagicDNS. It needs a public HTTPS host. The trade is one more server to run against everything in the table above.
Hermes: docs and plugins
Our Hermes Desktop clients already connect to the dashboard by LAN address, behind password sign-in. Over a WireGuard tunnel that routes the home LAN, the URL works unchanged. The Desktop docs describe a remote gateway as something to reach over "Tailscale or another trusted VPN". They say to keep a password-protected backend off the open internet and never forward its port from the router. A WireGuard tunnel meets that bar. Hermes' SSRF guard already treats RFC 1918 and 100.64.0.0/10 as private, so a 10.x tunnel subnet gets the same protection as the tailnet did.
| Catalog plugin | What it does | Fit here |
|---|---|---|
| wireguard | Read-only wg show status and peer list | Windows only. Not usable on our Macs or UniFi gateway, but it's a good pattern to copy. |
| netplane | Installs and runs a Headscale control plane, with join keys and a node list | Only for the Headscale middle path. Needs a fresh Debian/Ubuntu host. |
| hermes-tailscale | Desktop roster of tailnet machines | Retires with Tailscale. |
| hermes-vpn | Picks a VPN exit country for the agent's browser | Different problem. Not needed. |
No catalog plugin manages a UniFi WireGuard server. We plan to build two small, read-only pieces of our own. One is the DDNS-drift watchdog above. The other is a skill that reads handshake age per peer, so "is anyone's help tunnel up right now?" gets a real answer. Adding or removing peers stays a human decision in the UniFi UI, the same line the wireguard plugin draws.
Rollout
Tailscale stays up through every phase. Rollback at any step means using the tailnet as we do today.
| # | Step | Gate before moving on |
|---|---|---|
| 0 | Infra domain on Cloudflare, scoped token, gateway DDNS, LAN updater, watchdog cron | Name matches the public IP; watchdog fires on a forced mismatch |
| 1 | WireGuard server on the live WAN, tunnel subnet outside 192.168.x and 100.64/10 | Handshake from a phone on cellular |
| 2 | James's phone and laptop, on-demand away from home Wi-Fi, DNS = gateway | Hermes Desktop sign-in and SSH to the lab over cellular |
| 3 | Second site, site-to-site | Traffic both directions |
| 4 | First helped Mac: help server + zone, App Store tunnel, Get help / Done shortcut | Screen Sharing works while they have it on; nothing answers when it's off; they cannot reach our LAN |
| 4b | Collaborator servers, zones and host accounts, one person first | Their target works; every other LAN host, the gateway UI and the help zone time out from their tunnel |
| 6 | Two weeks with tailscale down and remote-desktop unattended access off, then log out and uninstall both | No path left that only Tailscale provides |
Still open
- Why we're leaving decides plain WireGuard versus Headscale. That's James's call, and this post doesn't assume one.
- Family phones and iPads that use the tailnet away from home. Phones roam across hotel and café Wi-Fi, which is where plain WireGuard is weakest. We need to know what they actually reach before choosing a design for them.
- Exactly which ports each collaborator group needs (SSH only, or also a model or API port) and whether anyone gets GPU time on the Station.
- Whether UniFi's site-to-site option fits our two consoles. We'll check that in the UI, not from a spec sheet.