Getting in

Everything here lives behind two doors. The first is Anubis, a quiet bot-check: the first time you visit any guarded page in a session you see a short "making sure you're not a bot" screen. It runs once. The second door is Authentik, the family single sign-on: log in once and every guarded service (this homepage, search, chat, media requests, and friends) opens for you without a second password.

• Need an account? Ask Frank. Accounts are issued by hand; there is no signup.
• Your clearance lives in a cookie. Clearing cookies means re-clearing both doors.
• If a page asks you to log in repeatedly, that is a server-side problem — tell Frank, don't brute-force it.

Talking to the AI

The everyday front door is Open WebUI (it has its own login the first time; use the same email you use everywhere). Pick a model in the chat's model selector — the three models the fleet runs are described below. Open WebUI keeps your chat history; the raw model endpoints do not.

There is no audio input anywhere on the fleet. Text in, and — on frank-small only — images too.

The running models

Three models serve the family, each on its own hardware. The homepage status cards above show whether each one is up right now.

frank-large — the big brain

• Model: Qwen3.8-Flash-Next, w4b quantization, served by halogen-flash-server.
• Hardware: Dankclaw's Strix Halo iGPU (gfx1151), shared with the whole host.
• Best for: long, thoughtful text conversations. 200k-token context per conversation.
• Capacity: exactly two conversations resident at once. A third waits — politely, but it waits.
• Not for: images. It cannot see them.

frank-small — the eyes

• Model: Gemma4-12B QAT (HauhauCS Balanced, Q4_K_M) with MTP + n-gram speculative decoding.
• Hardware: Dankclaw's RTX 3090.
• Best for: anything with a picture — screenshots, photos, diagrams. Fastest decode of the fleet at long context, and the endpoint that gets most daily traffic.
• Capacity: one million tokens of pooled context across four simultaneous conversations (up to 262k each).

frank-medium — the 27B generalist

• Model: Swift-Qwen3.8-27B-Uncensored at UD-Q8_K_XL, with its own MTP head, on mainline llama.cpp.
• Hardware: Dankserver's two Radeon R9700s, one process split 35:30 across the cards.
• Best for: general chat and reasoning when the big brain is busy; the biggest pure-language model we run.
• Capacity: 393k-token context over two conversations. Also sees images (BF16 projector).

Which one to pick? Screenshots and photos: frank-small. Deep long-text work: frank-large (mind the two-seat limit). When in doubt: frank-medium.

Media & files

• Movies & TV: request with Seerr; watch with Plex or Jellyfin. Don't ask the *arrs directly — Seerr is the front door, and requests ride the tunnel from there.
• Books: Calibre-Web — the e-book library, with its own login.
• Games: RomM — ROM library and emulator collection.
• Raw files: Copyparty — direct downloads and shares.
• Search: SearXNG — metasearch, VPN-routed, no AI snoop.

When something looks broken

• A status card says DOWN. Reload the page once; the checker re-probes every few minutes. If it stays down, that service really is unhappy — tell Frank.
• "status unavailable" on the homepage. The page couldn't read its status file. Hard-reload (Ctrl/Cmd-Shift-R).
• Connection timeout on a home device. Nine times out of ten the device isn't using the house DNS (Pi-hole at 192.168.0.143) and resolved our names to the public address, which doesn't answer. Point the device's DNS at 192.168.0.143 and retry.
• The bot-check spins forever. Let it finish; heavy JS blocking (privacy extensions) can starve it. Pause the blocker, reload, resume.