In this article
The position
Our position: for family-facing apps, a reverse proxy open to the whole internet is the wrong default. We think it is the largest attack surface many homelabs carry, and that is our judgement, not a measurement. We would accept the setup cost of an overlay network or an outbound tunnel to avoid running that listener at all.
What the evidence says
Tailscale Funnel sends traffic from the wider internet to a service on one of your tailnet devices, and visitors do not need Tailscale. Serve is the sibling feature, and it shares a service only with devices inside the tailnet. Funnel is off by default, and you turn it on through the Tailscale CLI.
Public DNS for a Funnel URL points at a relay server, not at your device. The relay opens a TCP proxy to your device over Tailscale, and TLS ends on your own node before the request reaches the local service. Tailscale says the relay cannot decrypt that traffic.
The vendor states that your device's IP address is never exposed to the internet through Funnel. A Funnel URL points at the one resource you share, such as a web server, a file or a directory, not at the whole device. Funnel still makes that resource reachable by anyone on the internet.
Serve and Funnel cannot share a port at the same time. If the latest command for a port was serve, the port is fully private, and if it was funnel, the port is fully public. That is a handy guardrail, and it is also an easy place to slip.
Funnel only works with DNS names in your tailnet's ts.net domain. It only listens on ports 443, 8443 and 10000, and it only accepts TLS-encrypted connections. Traffic over a Funnel is subject to bandwidth limits that you cannot configure.
It needs Tailscale v1.38.3 or later, MagicDNS, HTTPS with valid certificates, and a funnel node attribute in the tailnet policy file. Enabling Funnel from the CLI creates valid HTTPS certificates and updates the policy file for you. By default, any user in autogroup:member may use Funnel, and you can change that by editing the policy file.
Asking for new certificates too often can hit Let's Encrypt rate limits, and the docs say you may wait 34 hours before trying again. Serving a file, directory or Unix socket needs root, while forwarding a port does not. Cloudflare Tunnel connects resources to Cloudflare without a publicly routable IP.
The cloudflared daemon runs in your infrastructure and makes outbound-only connections to Cloudflare's network. It can carry HTTP web servers, SSH servers, remote desktops and other protocols. Cloudflare says you can set the firewall to allow only cloudflared's outbound connections and block all inbound traffic.
That blocks access to the origin from anything other than Cloudflare. According to Cloudflare, origins can serve traffic through it without being open to attacks that go around Cloudflare. A tunnel is a persistent object with a UUID, and you can run several cloudflared connectors in one tunnel.
Each connector sends traffic to the nearest Cloudflare data center, and you create tunnels in the dashboard or through the API. Put together, both products let you publish or reach a service without opening an inbound port on your router. Here is what the sources do not tell us, so you can weigh our verdict fairly.
None of them describes Caddy or NGINX, their defaults, or what a public listener exposes. None of them measures whether a public proxy is the largest attack surface in a typical homelab, so that claim rests on our reasoning. NetBird is not covered at all.
The Cloudflare page we used does not mention Cloudflare Access, so we cannot describe its login policies. We also have no documentation on how Jellyfin, Immich or Nextcloud mobile apps behave behind these paths, or on tunnel limits for large uploads or media. The Tailscale page text we had was cut off, so details after the certificate step of the funnel command are unknown to us.
The strongest objection
The best case against us starts with the people who use your services. Funnel lets visitors reach a service even if they do not use Tailscale, so there is no enrolment step for them. For a relative who struggles with new apps, skipping enrolment is a real advantage.
The objection continues that these paths bring limits of their own. Funnel is bound to three ports and to bandwidth limits you cannot change. Funnel traffic also crosses a relay server, and cloudflared traffic crosses Cloudflare's network, so a vendor sits in the path either way.
Funnel also depends on MagicDNS, HTTPS and a policy attribute being set up correctly. Running your own proxy gives you full control, and that control has real value.
No source we have measures how often public proxies are compromised, so the claim that the risk is small cannot be settled here. We still think the trade favours removing the listener, but we hold that view with that gap in mind.
Where it still makes sense
A public port is the honest design when the service is meant for the public, such as a website for strangers. It also fits when a workload will not fit inside a vendor path. Funnel serves only three ports and applies bandwidth limits you cannot adjust, so a workload that needs more than that may not fit.
If you patch quickly, watch your logs and keep few apps behind one proxy, a public port is defensible. Neither source says when a public port is the right call, so this verdict is our framing.
What to do
Here is a short order of work. First, list every app that answers on a public port today, and decide for each whether the public really needs it. For family apps, start private: a serve command leaves the port tailnet-only, so run tailscale serve for that port (verify before running).
If you must publish something, tailscale funnel is the command, and it asks you to approve the change in a web interface (verify before running). Check the requirements first: a recent client, MagicDNS, HTTPS certificates and the funnel attribute in the policy file. After enabling Funnel, review the policy file, because any autogroup:member may use it by default.
Re-check the last command you ran on each port, because it decides whether that port is private or public. Do not request certificates repeatedly while testing, since the docs mention a possible 34-hour wait. If you prefer Cloudflare, create the tunnel in the dashboard or through the API, then run cloudflared next to your service.
Once it works, change your firewall to allow only cloudflared's outbound connections and block inbound traffic, and test from outside your network (verify before running any firewall change). We cannot give Cloudflare Access steps, because our source does not cover them.
Enrol family devices one at a time and test each mobile app yourself, since we found no documentation on how they behave. If you keep a proxy on your network, keep it for local use and treat any public exposure as a deliberate, reviewed exception.
Sources
Image: Photo by Lightsaber Collection on Unsplash
This story was produced with AI assistance from the sources listed above and fact-checked against them before publication. Read our editorial methodology and the corrections log. Spotted an error? Email [email protected].





Discussion