The short answer
Tailscale is the default choice for home server remote access in 2026: free for personal use, zero port forwarding, and it connects from anywhere including CGNAT connections. Choose self-hosted WireGuard when you want no third party anywhere in the path and your home has a public IP. Choose ZeroTier when you need Layer 2 networking that broadcast-dependent tools require. None of these decisions is about raw speed — it is about trust model, setup effort, and whether your ISP gives you a reachable address.
Why this decision matters more than throughput
Remote access is the one feature that turns a home server from a toy into infrastructure. It is also the decision people postpone longest, because every option looks similar in a feature grid. What actually separates them:
- Who can see your metadata. A mesh VPN’s coordination server learns which devices exist and when they connect. Raw WireGuard has no coordination server at all — you are the only party that knows the topology.
- Whether inbound connections are possible at all. Behind CGNAT, a self-hosted VPN endpoint is unreachable no matter how fast it is. Mesh services traverse NAT because both peers make outbound connections to a common coordinator.
- Setup and maintenance cost. Mesh services exchange hours of key management for an account login. Raw WireGuard trades that back: every peer needs a keypair, an allowed-IPs block, and manual updates when the topology changes.
Throughput, the number comparison articles lead with, is nearly irrelevant on home internet. Your remote upload path is capped by your ISP’s uplink — typically a few tens of Mbps — while every option here can saturate that many times over on modest hardware. The measured numbers below exist to prove the hardware side of that claim, not to rank the VPNs.
The three architectures at a glance
| Tailscale | Self-hosted WireGuard | ZeroTier | |
|---|---|---|---|
| What it is | Managed mesh overlay on WireGuard | Raw VPN protocol (Linux kernel since 5.6, ~4,000-line codebase) | Managed L2/L3 virtual Ethernet |
| Coordination | Hosted controller (self-hostable via Headscale, unofficial) | None — you manage keys | Hosted root servers (controller is open source) |
| NAT traversal | Automatic, DERP relays as fallback | Manual — needs public IP + forwarded UDP port | Automatic via root servers |
| Network model | Layer 3 | Layer 3 | Layer 2 (broadcast/multicast work) |
| Free tier (verified Oct 2026) | 6 users, unlimited user devices, 50 tagged resources | Entirely free (self-hosted software, no service) | 10 devices, 1 network |
| Setup effort | Minutes | Hours, ongoing key management | Minutes |
| Trust requirement | Trusts Tailscale’s coordination plane | Trusts nobody | Trusts ZeroTier’s root servers |
One nuance people miss: Tailscale and raw WireGuard are not competitors at the protocol level — Tailscale connections are WireGuard tunnels. The competition is between managed coordination and self-managed keys. ZeroTier is the odd one out with its own protocol and the only option here that behaves like a virtual switch.
What we measured on the hardware all three run on
We have not benchmarked the VPN implementations themselves — this page states that plainly rather than borrowing numbers from other people’s routers. What we did measure, on the Beelink EQi12 (i3-1215U, 16 GB RAM) that terminates our own remote access, is the hardware ceiling any of them would operate against. Every value traces to a published raw file in the EQi12 measurement repository (CC BY 4.0).
| Measured on the EQi12 | Result | Measurement condition |
|---|---|---|
| Ethernet port 1, both directions | ~890.75 / ~902.57 Mbps | 4 GiB payload, controlled LAN |
| Ethernet port 2, both directions | ~912 / ~941 Mbps | 4 GiB payload, controlled LAN |
| SMB 50 GiB transfer | 107.15 / 102.95 MiB/s | SHA-256 verified both ways |
| Wi-Fi 6 (Intel AX200), near range | ~618 / ~870 Mbps | 4 GiB, 2.4 m from AP |
| System idle power | ~12-14 W | Wall meter, Docker stack running |
| Full load power | ~37 W | Wall meter |
| Wake-on-LAN from S5 | 5/5 successful | Cold shutdown, magic packet |
| 12.57-hour stability run | 0 failures | 150 samples, Docker stack |
The port-by-port methodology lives in the EQi12 network performance report. The practical reading: a 6-core mini PC idling at 12-14 W has orders of magnitude more packet-processing capacity than a home uplink needs, so VPN choice should be decided on trust, topology, and maintenance — never on throughput projections. The dual-port result also matters for one specific pattern below: a dedicated management NIC.
Decision paths
“I just want to reach my server from my phone and laptop.” Tailscale. Install on the server, install on your devices, log in on each. When your home connection is CGNAT — increasingly common on fiber and 5G fixed wireless — this is the only option that works without additional infrastructure. The Jellyfin remote access guide walks through exactly this path for media streaming.
“I will not put a third party in my network path.” Self-hosted WireGuard, with two prerequisites: a public IP on your home connection, and acceptance that you own key rotation, peer adds, and allowed-IPs bookkeeping forever. A common hardening pattern on dual-NIC mini PCs is terminating WireGuard on one port while services live on the other — our dual-ethernet WOL port selection guide covers the port-role question, including the measured 5/5 S5 wake reliability that makes remote wake viable before you even connect.
“I need devices to behave like they share a switch.” ZeroTier is the only option here with native Layer 2. Chromecast discovery, SLP, some printer protocols, and certain game server LAN advertisements rely on broadcast or multicast, which Layer 3 VPNs deliberately do not carry. The cost: the free plan now covers 10 devices and 1 network — reduced from an earlier 25-device free tier — so device-heavy labs hit its wall quickly.
“I want mesh convenience but no hosted control plane.” Headscale — an open-source reimplementation of Tailscale’s coordination server — runs on any small VPS or even the mini PC itself if you have a public address. You keep the Tailscale client experience, you own the coordinator, and you inherit its maintenance. This is the middle path most homelab guides skip.
“I expose services to the internet anyway.” Then ask whether you need a VPN for everything or only for the management plane. A reverse proxy in front of public-facing apps plus Tailscale for admin interfaces (router login, Proxmox, NAS admin) is the pattern the self-hosting alternative selector encodes as its default recommendation. Our nginx container benchmark (10 concurrent connections at 6,156 RPS, 0 errors, on the same EQi12) suggests the reverse-proxy side of that split is not a resource concern on this class of hardware.
The CGNAT reality check
This deserves its own section because it silently invalidates one option for a large share of households. Carrier-grade NAT means your router’s WAN address is shared; inbound UDP to your WireGuard port never arrives. Check your router’s WAN IP against a “what is my IP” service — if they differ, you are behind CGNAT. From that moment your realistic choices are: a mesh VPN (Tailscale or ZeroTier, both designed for exactly this), a small VPS as a WireGuard hop (~$4-5/month, which reintroduces a third party anyway), or asking your ISP for a public address. This is why our short answer leads with Tailscale: it is the only branch of the decision tree with no infrastructure prerequisite.
The exit node question
One capability check that feature grids usually flatten: routing all of a remote device’s traffic through your home server, not just traffic aimed at it. This is what makes a mini PC double as your travel router — hotel Wi-Fi, coffee-shop networks, and untrusted hotspots all exit through your home connection. Raw WireGuard handles this with an AllowedIPs = 0.0.0.0/0 peer entry, which works but hijacks the entire routing table and needs a split-DNS plan. Tailscale packages the same idea as a first-class “exit node” toggle you flip per device from its menu — pick the EQi12-class server as the exit node and done. ZeroTier routes per-managed-network only; pushing general internet traffic through it means manual route configuration on the client. If remote privacy is part of why you are reading this page, that single feature difference should weight your choice more than any performance column above.
Resource footprint on a mini PC
All three run comfortably on fanless-class hardware. Raw WireGuard is the lightest — an in-kernel module with no userspace daemon beyond the key manager. Tailscale’s daemon sits at tens of MB of RAM. ZeroTier’s service is comparable. None of this registers against a 16 GB home server already running Docker stacks; the honest constraint on a mini PC is not VPN overhead but the 512 GB single-NVMe storage budget and the 12-14 W idle wall draw we measured. If you are still choosing the machine itself, the Windows mini PC home server roadmap sequences remote access as a later stage precisely because it is cheap to add and easy to change later.
Fact block (for citation)
| Fact | Value | Source |
|---|---|---|
| Measured NIC throughput, EQi12 port 1 | ~890.75 / ~902.57 Mbps | EQi12 Measurement Data repo, 4 GiB LAN test, July 2026 |
| Measured NIC throughput, EQi12 port 2 | ~912 / ~941 Mbps | Same repo, same conditions |
| Idle / full-load wall power | ~12-14 W / ~37 W | Same repo, wall meter, 7-state series |
| WOL from S5 | 5/5 successful | Same repo, cold shutdown wake series |
| Tailscale Personal free tier | 6 users, unlimited user devices, 50 tagged resources | tailscale.com/pricing, verified October 2026 |
| ZeroTier Personal free tier | 10 devices, 1 network | zerotier.com pricing, verified October 2026 |
Data licensed CC BY 4.0. VPN product terms change; verify current quotas before making purchase decisions. This page does not contain VPN throughput benchmarks — the tools themselves were not benchmarked by us.
Division of labor with the rest of this site
- EQi12 network performance report — the raw NIC and Wi-Fi measurements cited above.
- Jellyfin remote access setup — the configuration walkthrough once you have picked a VPN.
- Self-hosting alternative selector — a quick selector across all self-hosting categories; this page is the deep dive on its remote-access row.
- Self-hosted password manager on a mini PC — the service that pairs most naturally with a VPN-only front end: Vaultwarden sized on the same measured hardware.
- Dual-ethernet WOL port selection — which NIC should face the internet vs the LAN.