Services

Ports

Nothing listens on the LAN except Jellyfin and mDNS. Everything else is loopback, the tailnet, or inside the VPN namespace.

Service	Address	Reachable from
Jellyfin	:8096 http, :8920 https	LAN (enp5s0) and tailnet (tailscale0)
qBittorrent web UI	:8081 in namespace wg	tailnet via port mapping; host at http://192.168.15.1:8081
Prowlarr	:9696 in namespace wg	tailnet via port mapping; host at the namespace address
FlareSolverr	127.0.0.1:8191 in namespace wg	Prowlarr only
Radarr / Sonarr	:7878 / :8989	host only (firewall closed)
Audiobookshelf	127.0.0.1:8000	host
calibre-web	127.0.0.1:8083	host
llama.cpp units	127.0.0.1:8080	host
llama-fim	127.0.0.1:8012	host
Ollama	127.0.0.1:11434	host
MPD	127.0.0.1:6600 (socket-activated)	host
hello-page	127.0.0.1:8137 → tailscale serve on 443	tailnet (HTTPS)
Avahi	:5353/udp	LAN
claude-vm	127.0.0.1:2222 (ssh), :5173, :3000	host, while the VM runs

Media

/mnt/media (modules/media.nix) mounted with the user as owner and gid 3000 (media), dmask=0002 / fmask=0113 so the whole media group can write. It's nofail + x-systemd.automount; every unit that touches it has RequiresMountsFor = /mnt/media, so services wait for the drive instead of writing into an empty mountpoint.

Service	Notes
Jellyfin	NVENC. Decode: h264, hevc, hevc10, vp8, vp9, av1, mpeg2, vc1. Encode: HEVC on, AV1 off. Tone mapping on, transcode throttling on, CRF 21 (h264) / 26 (h265). DeviceAllow adds nvidia0, nvidia-uvm(-tools), nvidia-modeset. forceEncodingConfig = true — encoding settings changed in the web UI are overwritten on restart; change them in Nix
Radarr / Sonarr	group media, lib/hardening.nix with ReadWritePaths = /mnt/media
Audiobookshelf	loopback only, group media, hardened
calibre-web	loopback only, library at /mnt/media/books, uploads disabled. GOOGLE_BOOKS_API_KEY (optional) comes from secretspec
MPD	user service (home/profiles/linux-desktop.nix), music at /mnt/media/music. Two outputs: PipeWire (default) and a bit-perfect ALSA output on hw:CARD=G30 (disabled; switch with mpc enable only 2, back with mpc enable only 1). Clients: rmpc, mpdris2 for media keys

VPN namespace

modules/vpn.nix, using VPN-Confinement.

            tailnet 100.64.0.0/10
                    │  port mappings 8081, 9696
host ─── bridge ────┤
                    ▼
      ┌─ netns "wg" ── WireGuard (Mullvad) ── internet
      │   qbittorrent  :8081   torrent port 51413
      │   prowlarr     :9696
      │   flaresolverr 127.0.0.1:8191
      └─

• The namespace is built from /etc/wireguard/mullvad.conf (not in the repo). wg-dns runs before wg.service and rewrites the file's DNS = line to Mullvad's resolver 100.64.0.7, so lookups inside the namespace never leave the tunnel.
• accessibleFrom = [ tailnet ] routes replies to tailnet clients back out of the bridge instead of into the tunnel.
• A firewall rule drops forwarded traffic to 8081/9696 unless it came in on tailscale0 — the mappings are tailnet-only, not LAN.
• qBittorrent runs as its own system user in group media, saves to /mnt/media/.incoming, no UPnP, pauses at ratio 2 or after 24 h seeding, MemoryHigh = 2G. Web UI auth is skipped for the namespace subnet (192.168.15.0/24); the host reaches it at http://192.168.15.1:8081, which is what the bar's dots-qbt widget polls.
• Prowlarr is hardened with lib/hardening.nix.
• FlareSolverr gets a browser-friendly variant of the hardening (Chromium needs namespaces and devices) and waits for DNS inside the tunnel before starting.

Check confinement from inside the namespace:

sudo ip netns exec wg curl -s https://am.i.mullvad.net/connected

Mullvad for the host itself is separate (modules/mullvad.nix, the GUI and dots-vpn toggle).

Local LLMs

modules/llm.nix + modules/ollama.nix. Models live on /data/models.

Unit	Model	Port	For
llama-cpp	Qwen3.5-9B Q6_K	8080	general / coding / maths
llama-sec	WhiteRabbitNeo V3 7B Q5_K_M	8080	security
llama-agent	Qwen2.5-Coder-14B-Instruct Q4_K_M	8080	tool calling
llama-gemma	Gemma 4 12B Q4_K_M	8080	general, vision, audio
llama-coder	Gemma 4 26B-A4B MoE Q4_K_M	8080	coding / agentic — 10 expert layers on CPU, q8_0 KV cache
llama-fim	Qwen2.5-Coder-3B Q8_0	8012	llama.vim fill-in-the-middle
ollama	/data/models/ollama	11434	trying models quickly

How they behave:

• Vulkan build of llama.cpp (llama-cpp-vulkan) — cached, unlike the CUDA build. All layers on GPU (-ngl 99), 16k context, flash attention, --jinja chat templates.
• One at a time. Every llama unit Conflicts= every other llama unit, so starting one stops the last. Before starting, free-gpu asks Ollama to unload whatever it holds. Ollama keeps at most one model loaded and drops it after 5 minutes idle.
• Never at boot. wantedBy is forced empty on all of them, Ollama included.
• No sudo. A polkit rule lets your active local session start and stop these units.
• Sandboxed. The llama units run as DynamicUser with ProtectHome, ProtectSystem=strict, no capabilities and MemoryDenyWriteExecute. GGUF files must be world-readable.

systemctl start llama-coder      # swap models
dots-llm next                    # cycle, same as the bar widget
dots-llm off
gpu-check                        # which llama unit is up + ollama ps
gpu-free                         # stop all llama units

Backups

modules/backup.nix — restic, $HOME → /mnt/backup/restic, daily (persistent, up to 1 h random delay).

• Excludes caches, Steam, flatpak and container storage, Downloads, Trash, and build junk (node_modules, target, zig-cache, zig-out, .venv, .direnv, __pycache__).
• Retention: 7 daily, 4 weekly, 6 monthly. Each run checks a random 5 % of pack data.
• Skips cleanly when the backup drive isn't mounted; runs at nice 19, idle I/O.
• Password at /etc/restic/password; the unit refuses to start if the file is empty.

sudo systemctl start restic-backups-home
journalctl -u restic-backups-home -e
sudo restic-home snapshots        # wrapper with repo + password preset

Storage

Mount	What	Mounted as
/	LUKS, ext4
/data	LUKS2 ext4, unlocked after root by keyfile (crypttab), nofail	games, models, VMs, replays
/mnt/media	exFAT 10 TB	user:media, group-writable, automount, visible in file managers
/mnt/backup	exFAT	root:root 0077, automount (10 min idle), hidden from file managers
/mnt/newvolume	exFAT	user:users, automount (10 min idle)

exFAT has no permissions, so ownership comes from mount options. udiskie ignores all three external drives by UUID — they belong to systemd — but still automounts real removable media. dots-mounts (the bar's storage alert) reports any of /data /mnt/media /mnt/backup /mnt/newvolume whose mount or automount unit isn't healthy.

/data directories are created by tmpfiles rules in the modules that use them: /data/games (gaming), /data/models and /data/models/ollama (LLM), /data/vms and /data/vms/iso (virt), /data/replays (recording).

Flatpak

modules/flatpak.nix declares remotes and apps; the flatpak-managed user unit reconciles at login.

• Remotes: Flathub and NVIDIA's GeForce NOW repo.
• Apps: GeForce NOW, Foliate, Flatseal, Bottles, F3D, MeshLab, Blender, Simple Scan, Kdenlive, TeXstudio.
• It installs anything listed and uninstalls any user app that isn't — add apps here, not with flatpak install.
• GeForce NOW gets SDL_VIDEODRIVER=x11. The desktop profile exports SDL_VIDEODRIVER=wayland globally, flatpak inherits it, and SDL2's Wayland backend can't capture the mouse — aim breaks in-game without the override.

hello-page

modules/hello-page.nix runs ~/the-page/app.py (system python, as your user, heavily sandboxed, read-write only to that directory) on 127.0.0.1:8137. hello-page-serve waits for tailscaled, then publishes it with tailscale serve --bg --https 443. Environment overrides go in /var/lib/secrets/the-page.env.

Containers and VMs

• Docker runs rootless; DOCKER_HOST points at the user socket. Not started at boot.
• Podman is available alongside.
• libvirt: qemu:///system, images under /data/vms; virt-manager is preconfigured (autoconnect to system, SPICE, host-passthrough CPU, qcow2). The agent VM is separate — see Development.

---

_{Source: docs/services.md}