Compare / Self-hosting decision

Traefik vs Nginx Proxy Manager vs Caddy on a Mini PC: Ports, Certificates and an Honest Comparison

Which reverse proxy fits an already-busy home server? The documented port map, how each one handles TLS certificates, and where automatic HTTPS does and does not remove work — verified November 2026, with our measured EQi12 host data kept separate from application claims we did not test.

The short answer

Choose on configuration model, port fit and certificate workflow — not on a performance claim. We have not benchmarked any of the three, and none of them publishes a number that would let you rank them on speed. What you can verify before installing is more useful anyway: which inbound ports each one wants, how it decides where traffic goes, and how much of the certificate lifecycle it handles for you.

Nginx Proxy Manager is a prebuilt Docker image with a web admin interface on port 81; you add proxy hosts in a UI and it obtains Let’s Encrypt certificates for you. Traefik is a routing engine configured by providers — in Docker, by labels on the containers themselves — with TLS configured as a separate documented step. Caddy is a single binary configured by a Caddyfile or a JSON API, and it is the only one of the three whose official documentation says HTTPS is automatic and on by default, including the HTTP-to-HTTPS redirect and certificate renewal.

On a mini PC that already runs Windows, Docker Desktop and a media stack, the first real decision is almost always who owns ports 80 and 443 — not which proxy has the cleverest feature list. A busy host has opinions about that. Ours had 25 unique listening TCP ports before we installed anything else.

What we measured, and what we did not

This site publishes original measurements of one specific machine: a Beelink EQi12 (Intel Core i3-1215U, 16 GB DDR4, WD SN540 512 GB, two Realtek gigabit ports, Intel AX200 Wi-Fi 6). Every hardware number below comes from that machine and traces to a file in the public measurement repository (CC BY 4.0).

We have not installed, run, timed or load-tested Traefik, Nginx Proxy Manager or Caddy on that machine. All three applications are therefore in our “not tested” category, and we will not quote a requests-per-second figure, a memory footprint, a latency number or a TLS handshake cost for any of them. Where a number appears in this article it is either (a) a vendor-documented port, configuration default or requirement, clearly labelled as such, or (b) an original measurement of the host, with the label saying so. Arithmetic done on measured host values is marked as derived arithmetic, not a measurement.

That separation is the whole point of the page. A hardware ceiling is not an application benchmark, and confusing the two is how “my 16 GB mini PC can proxy a million requests a second” claims get made. Our testing methodology explains the five evidence levels we use and why we keep failed runs in the record.

Evidence 1 — the documented configuration, side by side

These rows come from each project’s own published documentation, checked on 11 November 2026. Nothing in this table is our measurement.

Documented propertyNginx Proxy ManagerTraefikCaddySource
Published version in the guidejc21/nginx-proxy-manager:2.16.0 imagetraefik:v3.7 imageCurrent docs (no pinned version in the getting-started guide)NPM quick setup; Traefik Docker quickstart; Caddy getting started
Documented web ports80:80 and 443:44380:80 for the web entrypoint; TLS is a separate “Configure TLS” stepHTTP on default port 80, HTTPS on default port 443, redirect automaticNPM guide; Traefik docs; Caddy Automatic HTTPS
Admin / dashboard port81:81, admin UI at http://127.0.0.1:818080:8080 dashboard, only when --api.insecure=true (docs: development use only, do not enable in production)Admin API at localhost:2019, restricted to localhost by defaultNPM guide; Traefik Docker quickstart; Caddy getting started
Primary config modelWeb admin interface: proxy hosts, redirections, streams, 404 hosts, access listsProviders; in Docker, labels such as traefik.http.routers.whoami.rule=Host(...)Caddyfile (adapted to JSON) or native JSON via the APINPM features list; Traefik Docker quickstart; Caddy getting started
Automatic service discoveryNo — hosts are added in the UIYes — the Docker provider reads container labels; requires mounting /var/run/docker.sockNo — routes are declared in configTraefik Docker quickstart; NPM guide; Caddy docs
TLS certificatesFree SSL via Let’s Encrypt, or your own certificatesConfigured as a separate documented stepAutomatic HTTPS on by default; certificates obtained and renewed, HTTP redirected to HTTPSNPM guide; Traefik next steps; Caddy Automatic HTTPS
Default certificate authoritiesLet’s Encrypt (or custom)Configured by the operatorLet’s Encrypt and ZeroSSL, with automatic fallback if the first failsNPM guide; Caddy Automatic HTTPS
Local / internal HTTPSBring your own or use a custom certificateConfigured by the operatorGenerates its own local CA and signs certificates for localhost, 127.0.0.1 and internal names; clients must trust that rootCaddy Automatic HTTPS
Config reload behaviourManaged through the UIProvider-driven, dynamicZero-downtime config reload via caddy reload or the API; rolls back to the last working config on errorCaddy getting started
Prerequisites statedDocker and Docker ComposeDocker (Compose optional)caddy and curl on PATH for the tutorial; install by package manager or binaryNPM quick setup; Traefik quickstart; Caddy getting started

Two things are worth pulling out. First, “automatic HTTPS” is a Caddy default, not an industry default: in the Traefik quickstart, TLS is explicitly a next step, and in NPM it is a feature you enable per proxy host. Second, the three config models imply different daily work — a UI, labels on each container, or a text file — and that is a much better basis for a decision than a feature checklist. A vendor-authored feature table describes the configuration the author had in mind, not your stack.

Evidence 2 — the port map against a real Windows + Docker host

All three projects publish port lists. What they cannot publish is the state of your host. We audited a real one: a Windows 11 + Docker Desktop EQi12 with Jellyfin, PostgreSQL, nginx and Redis running. That audit found 25 unique listening TCP ports across 36 listener bindings and 169 enabled inbound allow rules in Windows Defender Firewall. The full inventory, including the exact PowerShell commands, is in the Windows home server port audit.

Mapping the documented proxy ports onto that measured environment gives a concrete pre-install checklist:

Port / protocolWho wants itCollision risk on our audited hostWhat to check first
80 TCPAll three (HTTP, redirects, ACME HTTP challenge)Not among the 25 measured listenersAny existing web server or published container port; router port-forward
443 TCPAll three (HTTPS)Not among the 25 measured listenersSame, plus whether you already terminate TLS elsewhere
81 TCPNPM admin UINot occupied in our inventoryNever expose it publicly; it is the control plane
8080 TCPTraefik insecure dashboard (dev only)Not occupied in our inventoryDocs say do not use --api.insecure=true in production
2019 TCPCaddy admin API (localhost-only by default)Not occupied in our inventoryKeep it bound to localhost; it can change config without a file
8096 TCPJellyfin (already running)Occupied, bound to 0.0.0.0 in our auditThis is a backend you forward to, never a port the proxy should own
139 / 445 / 2179SMB and Hyper-V managementOccupied, Windows-ownedNothing to do with the proxy — do not unbind them
18080 / 18096Docker Desktop backendOccupied, internalLeave alone; Docker Desktop manages these

The practical rule: the two web ports are usually free on a Windows home server, but the binding scope is where mistakes happen. Our audit found SMB’s 445 bound only to :: while 139 was bound to four specific addresses including the Docker bridge interfaces, and Jellyfin published on 0.0.0.0:8096. Docker Desktop publishes container ports on every interface by default, which is why a service you thought was local can already be reachable from the LAN. Before you hand 80/443 to a proxy, decide whether the admin port (81, 8080 or 2019) is reachable from outside — it should not be. Our port audit checklist generator produces the ordered before/after diff for exactly this change.

Evidence 3 — host headroom, and the arithmetic we will not overclaim

We can tell you what the host looked like under a real four-container workload. We cannot tell you what any of these three proxies would add to it, because we never ran them.

Host propertyMeasured valueWhere it came from
Wall power, Windows idle14 WPower consumption lab
Wall power, four healthy containers12 WSame lab; also in the 135-measurement readout
Wall power, full CPU load37 WSame lab
Container memory320 MiB across four containers, against a 7.638 GiB ceiling16 GB RAM budget walkthrough
Windows-side memory10,232 MB during the same runSame walkthrough
Wired throughput891–941 Mbps across both physical ports, both directionsEQi12 network performance report
Sustained stability run12.57 hours, 150 samples, zero HTTP failures, zero container restarts, max sampled CPU 52%12.57-hour stability test

Read that table as a description of the platform, not a prediction. The 7.638 GiB container ceiling is a real measurement of our Docker environment; subtracting 320 MiB from it is derived arithmetic, not a measured headroom figure for a proxy process. A reverse proxy’s actual footprint depends on its config, TLS session volume and the number of routes — none of which we measured here.

One measurement-method caveat applies to any monitoring you do afterwards: on this same host, calling docker stats inside a tight sampling loop inflated measured idle CPU by 29.7% and stretched the loop itself from 5 s to 9.2 s. See the docker stats monitoring overhead test. The observer changes the number, which is exactly why we are not going to invent one for software we never installed.

Where the three projects genuinely differ

Configuration model is the real dividing line. Nginx Proxy Manager is built for someone who wants to click through a form: its own guide says the project’s goal is to make reverse proxying “so easy that a monkey could do it”, and its documented feature list includes access lists, basic HTTP authentication, user management with permissions and an audit log. That is a good fit when several people touch the box and you want a record of who changed a route.

Traefik is built for a stack that changes itself. Its Docker quickstart enables the Docker provider and mounts the Docker socket so a container with a traefik.http.routers.whoami.rule label is routed without anyone editing a central file. That is a strong fit for a Compose-heavy host where services come and go — and a real trust decision, because the socket is a privileged handle on the host. The same quickstart is explicit that --api.insecure=true is for development only. Our Docker Compose home server stack shows the kind of multi-service layout where label-driven discovery stops being theoretical.

Caddy is built for a text file and a default that does the right thing. Its documentation says automatic HTTPS is activated when the config contains a hostname or IP, that certificates are obtained and renewed for qualifying names, and that HTTP on port 80 is redirected to HTTPS on port 443. It also documents a local CA for localhost and internal names, so an internal-only service gets HTTPS too — with the honest caveat that any client not trusting Caddy’s root will show security errors. Its caddy reload performs zero-downtime config changes and rolls back to the last working config when a new one fails.

Certificates are where marketing language and operational reality separate. Caddy’s “it just works” claim has documented preconditions: DNS records pointing at your server, ports 80 and 443 open externally, Caddy able to bind or receive those ports, a writable persistent data directory, and the domain name present in the config. Miss one and the automation cannot complete its ACME challenge. The other two make certificate issuance an explicit step, which is more work up front but leaves fewer hidden assumptions about your router. None of the three removes the need to decide whether your services should be reachable from the internet at all — see the Tailscale vs WireGuard vs ZeroTier comparison for the case where a private network is the better answer.

Decision table

Your situationLeaningWhy, from documented facts
You want the fewest certificate decisionsCaddyAutomatic HTTPS on by default, with HTTP→HTTPS redirect and renewal documented
You want a web UI and an audit log of route changesNginx Proxy ManagerAdmin UI on port 81; user management, permissions and audit log in its documented feature list
Your Compose stack changes often and you want routes to follow containersTraefikDocker provider discovers services from container labels
You do not want to mount the Docker socketNPM or CaddyNeither documented basic setup requires it; Traefik’s Docker provider does
Ports 80/443 are already owned by something elseAny, with planningAll three want the same two web ports; one edge proxy is the workable pattern
You need HTTPS on internal-only names without a public domainCaddy, with a caveatIt generates a local CA; clients must trust that root or they will see errors
Your host is Windows 11 + Docker DesktopAny of the three in DockerOur measured host runs that combination; see the Windows build and the WSL2 and NAT note for the port-forwarding caveat
You are adding a small service to an already measured boxAnyThe self-hosted password manager sizing guide uses the same “documented requirement against measured host ceiling” method
You are still deciding on the OS underneath—The Windows + Docker vs Proxmox comparison covers the layer below the proxy

Deployment checklist on an already-busy host

  1. Inventory listeners before installing. Record TCP and UDP bindings and their owning processes. Our port audit publishes the exact commands and a real 25-port result to compare against, including the 0.0.0.0 versus 127.0.0.1 distinction.
  2. Decide who owns 80 and 443. Only one process can. If a DNS sinkhole or another web service already claims them, read the AdGuard Home vs Pi-hole port map for how a second service competes for the same sockets.
  3. Keep the admin interface off the internet. Port 81 (NPM), 8080 (Traefik insecure mode — development only per its docs) and 2019 (Caddy, localhost-only by default) are control planes, not public endpoints.
  4. Read the Compose file before you run it. Specifically, check whether it mounts /var/run/docker.sock. Traefik’s label discovery needs that mount; the other two documented setups do not.
  5. Verify the certificate preconditions, not just the feature. For public names: DNS A/AAAA records, open ports 80 and 443, a writable persistent data directory. Caddy documents all four explicitly; the other two need them too for Let’s Encrypt HTTP validation.
  6. Forward one backend first and test it. Our Jellyfin setup guide documents a real service on port 8096 that makes a sensible first backend, and the full build roadmap shows where the proxy fits in the sequence.
  7. Measure afterwards, with a light touch. Record your own listener diff and throughput after deployment. If you poll docker stats in a tight loop, remember the 29.7% observer effect we measured on this exact host.

The numbers and where they come from

ClaimValueSource
NPM published ports80, 81 (admin), 443Nginx Proxy Manager official quick setup, checked 2026-11-11
NPM image in the guidejc21/nginx-proxy-manager:2.16.0Same document
Traefik quickstart ports80 (web entrypoint), 8080 (insecure dashboard, dev only)Traefik official Docker quickstart, traefik:v3.7, checked 2026-11-11
Traefik label discoverytraefik.http.routers.* labels + /var/run/docker.sock mountSame document
Caddy portsHTTP 80, HTTPS 443, admin API localhost:2019Caddy getting started + Automatic HTTPS docs, checked 2026-11-11
Caddy automatic HTTPSOn by default; HTTP challenge needs 80, TLS-ALPN needs 443; Let’s Encrypt + ZeroSSL fallbackCaddy Automatic HTTPS documentation
Our host’s listening ports25 unique TCP / 36 bindings / 169 inbound allow rulesOur original port audit, EQi12_Security_02_TCP_LISTENERS.csv
Our host’s power, memory, network and stability values12/14/37 W, 320 MiB / 7.638 GiB, 891–941 Mbps, 12.57 h zero failuresLinked lab pages above, all traced to raw files in the measurement repository
Application performance, memory, latency for all three proxiesNot measured — no figure published hereThis page’s methodology

Where to go next

If you are choosing the next service to put behind that proxy, the self-hosted password manager sizing guide and the Homarr vs Homepage vs Dashy dashboard comparison apply the same “documented facts against measured host” method. For a service that competes for the same web ports rather than sitting behind them, read the AdGuard Home vs Pi-hole decision. For the hardware under all of it, the EQi12 review is the canonical specification record, and the self-hosting alternative selector narrows options from your RAM, storage and skill constraints.