🔒 Admin Settings

Facts Alfred remembers (say "remember that ..." or add below). Voided notes stay visible in history but are no longer recalled.

Add

Device names and rooms are assigned here — devices themselves can't rename. New devices (browsers, panels, satellites) appear in the ⏳ pending section on first connect. "Copy link" puts a claim URL on the clipboard — open it in a new device's browser to bind it to that row.

The house theme is the default for every display that hasn't picked its own day/night presets in Display settings. Device choices always win.

Preset (device default)

Accent color (--primary override)

Every alarm across devices — one-shot and recurring, including alarms set by voice or on a panel. Deleting here stops them everywhere. One-shot alarms ring on the device they target (or the server default when no device is picked).

(no device — server default) Chime Bells Music box Beep Rooster + Set alarm

One-shot for the next occurrence of that time (today, or tomorrow if it already passed).

Lists are managed here (admin) — household members fill them in from the dashboard, planner, or voice. Shopping and To-Do are permanent built-ins — clear only. Deleting is local-only: removed items are tombstoned and won't come back from Home Assistant or Google, but a remote copy keeps its old name — new items added there will re-create the list.

+ New List

Daily backup time 💾 Back up now

Off-site copy Test Save

Off-site copy is off — backups stay in data\backups on this machine only.

Daily backups, kept 14 days. Restore replaces ALL backed-up data — notes, lists, calendar events, automations, settings, patterns, voice profiles, alarms, reminders, webhooks, Hindsight memory (🧠) — with that backup's contents, then restarts the server (~1 min). A safety snapshot of the current state is written first, so a restore can itself be undone.

⟲ Undo last restore

Master index of every data repository — what information exists, where it lives, how fresh it is. Generated on demand; nothing here is stored.

Manage the RSS feeds Alfred reads (voice, briefings, news tab) and the topic widgets shown on the dashboard.

📺 News tab tip: widgets render in a 5-column grid — the first 10 (2 rows) fit the screen at once; anything beyond that needs scrolling. For fullscreen/kiosk dashboards, keep the Widget toggle on at most ~10 feeds + topics combined.

My Feeds

Loading master list… (refresh the inventory sites below) ＋ Add from List

Add

Topics & Interests

Topic widgets filter headlines across your feeds by keyword. Order = dashboard priority (top first).

Add

RSS Master Lists

Inventory sites that list available RSS feeds — they populate the dropdown above.

Add

News Settings

Fallback feeds — used only while My Feeds above is empty:

💾 Save Settings

All Levels Debug Info Warning Error ↻ Refresh

Patterns Stats Coverage Batch Test Import/Export Generate Priority A/B Test Record Map

50 100 200 500 ↻

50 100 200 500 ↻

Speaker Name

🎤 Record (5 sec)

Scene Name

📸 Capture Current State Apply Template... Apply + Add Step ▶ Test Scene 💾 Save as Command

+ New Webhook ↻ Refresh

Directory 📁 Set 🔍 Scan Folder

⬇ From URL

⬆ Upload File ↻ Refresh

🔔 Ringtone volume 100% Save

Library

Browse disk — add files / import folders

Browse 📥 Import 🏠 🎵 ⬇

Search the web

All audio Live Music Netlabels 78 RPM Jamendo 🔍 Search

Loading…

Live resources

Name	Type	State / Workers	Registry alias	Actions
Loading…

ℹ️ What do these mean?

What the metrics mean

LOAD — time to get the model into GPU memory. The first benchmark after an Ollama restart is "cold" (slowest, 5-15s). Subsequent runs are warm (<1-2s).

PROMPT (tok/s) — how fast the model reads your input text. Matters when commands carry long context. Higher is better; most modern models exceed 300 on this GPU.

GEN (tok/s) — generation speed for a short answer. This is the "reading speed" you experience in conversation.

SUSTAINED (tok/s) — generation speed over a 200-token answer. The most representative number — use this to compare models.

VRAM — GPU memory the model occupies while loaded. Pinned models (keep_alive) hold this permanently.

GPU FREE — headroom. Voice TTS + STT need ~1-2 GB; below 2 GB free, Ollama may evict models mid-conversation.

Rules of thumb:
• Green ≥ 30 tok/s · Orange ≥ 10 · Red < 10 — red usually means part of the model fell back to CPU (check GPU Memory above for pressure)
• Compare models by SUSTAINED at the same ctx value
• Instant commands (lights, timers, weather) never touch the LLM — the pattern matcher handles them in <100 ms, so model speed only affects "thinking" answers

Benchmark Runner

Exercises the real production paths (STT, wake word, LLM provider, router, sentence matcher, TTS, joke pool, database) with canned inputs. Runs automatically ~10s after every server start — the button below re-runs it on demand (~5-8s).

▶ Run Tests

Loading…

Runs the command catalog — every supported command shape through the REAL pipeline (wake strip → clause split → gates → patterns → router → handlers) with a mocked home + LLM, in an isolated subprocess. Asserts routing, instant-path coverage (commands that should NOT call the LLM), expected service calls, and response content. Takes ~10-20s.

▶ Run Command Test

📊 Usage 📄 Export PDF

Loading…