Alfred for Windows

A voice assistant for your PC that runs on the PC itself: no cloud AI. Say "Whiskey, …" or hold Ctrl+Alt+Space and speak. The assistant is called Alfred; its wake word is "Whiskey". A local language model works out what you meant, and Alfred does it:

• opens apps and websites;
• controls Spotify;
• sets timers and alarms;
• finds your files;
• reads your calendar;
• and more.

It's the Windows edition of the Mac notch assistant. The design is in Notch Assistant Windows Specification.md.

Status: everything below is built and unit-tested. The Windows app has been compiled but not yet run on a real Windows PC. Your first run is its first run, so please send back what you see (see Reporting problems).

What you need

• Windows 11 (24H2 or later), 64-bit.

• The .NET 10 SDK: download the x64 installer from https://dotnet.microsoft.com/download/dotnet/10.0.

• VS Code with the C# Dev Kit extension. VS Code offers to install it when you open the folder.

• For GPU speed (strongly recommended on the Omen):

  • an up-to-date NVIDIA driver;
  • the CUDA 12 toolkit: open https://developer.nvidia.com/cuda-toolkit-archive, pick the newest 12.x release, then Windows › x86_64 › 11 › exe (local). The default install is fine.

  Without CUDA, Alfred still works on the CPU, just more slowly (plans take 2–4 s instead of well under 1 s).

• About 4 GB of disk for the models, which Alfred downloads for you.

Run it

1. Clone the repo and open the folder in VS Code.
2. Press F5 (or Run › Start Debugging › "Alfred (Windows)"). You can also use a terminal:

   dotnet run --project src\Alfred.Windows
   

3. The first time, Settings › Model opens. Click Download everything needed (about 3.2 GB) and wait until it says "All downloaded". Each file is checked against its SHA-256 before it's used.
4. Alfred's icon, a white "A" on black, is in the notification area at the bottom right of the taskbar. You may need the ^ arrow to see it. Right-click it for the menu.
5. Hold Ctrl+Alt+Space, say "open Spotify", and let go.

To use the wake word, turn it on in Settings › General. It then listens for "Whiskey, …" (for example "Whiskey, open Spotify") while plugged in, and on battery too if you allow it.

What to say

Other ways to control it:

• Escape cancels anything.
• A ringing alarm stops when you say "Whiskey, stop" or "Whiskey, snooze", or when you click the buttons.
• Type a command… in the tray menu works without a microphone.

Safety rules:

• Alfred never runs commands and never reads what's inside your files.
• Deleting only ever goes to the Recycle Bin, and "undo that" reverses the last file change.
• It won't act on anything you didn't say.

Optional accounts

Each of these uses your own free developer credentials, entered in Settings. The repo contains none.

• Spotify songs and playlists by name: Settings › Spotify has the steps. Play, pause and skip work without any setup. Starting a particular song through Spotify's API needs Premium; without Premium, Alfred opens the song in the app and presses play.
• Calendar and reminders: Outlook or Google, in Settings › Calendar.
  • Alfred only reads the calendar. Reminders go into the same account's task list.
  • Without an account, reminders ring in Alfred.
  • Google asks you to sign in again every 7 days while your Google app is in "Testing".

Where things are

Reporting problems

Please send:

• the day's log file from %LOCALAPPDATA%\Alfred\logs;
• what you said;
• what happened.

The log records:

• whether the language model and Whisper loaded on the GPU. Look for "on the GPU" and "runtime Cuda"; "CPU" means CUDA isn't being used;
• every command heard and how long it took;
• any errors.

Things worth checking on the Omen:

• the pill appears at the top centre of the laptop screen, including with an external monitor plugged in;
• the pill stays hidden during a full-screen game, with results spoken instead;
• the first word isn't cut off when you hold the hotkey;
• "Whiskey" wakes it without false wakes from videos, calls or words like "whisper";
• a timer and an alarm ring, including with Alfred closed (a toast);
• rename a file, then "undo that".

For developers

dotnet build Alfred.slnx
dotnet test Alfred.slnx

Projects:

Speech tests:

• They need the models in models/.
• They also need test clips: on a Mac, run say -v Daniel -o x.aiff "Alfred, open Spotify." and convert with afconvert -f WAVE -d LEI16@16000 -c 1. See tests/Alfred.Speech.Tests.
• Without the models or clips they pass trivially.

Intent corpus: this runs the model against shared/fixtures/intents.txt. Nothing is started.

dotnet run --project src/Alfred.PlanCli -c Release -- --model models/Qwen3-4B-Q4_K_M.gguf < shared/fixtures/intents.txt

Results from Sep 27, 2026, for Qwen3-4B Q4_K_M on an M4 Mac with Metal:

• Every command plans correctly, or reasonably declines.
• Direct matches take under 10 ms.
• Model calls take a median of 2.3 s. The Omen's RTX should be several times faster.

Not in this version

• Hand gestures: the Mac has them; on Windows they're planned for later (spec §10).
• Kokoro voice: Alfred uses Windows' own voices. More natural ones are in Windows Settings › Speech. Kokoro's licensing question is still open (spec §13).
• YouTube autoplay: Alfred opens YouTube's results. Clicking the first video for you (the Mac's "Tier 2") isn't ported.
• Notes: they go to Markdown files in Documents\Alfred Notes, not to OneNote.
• Full screen: "full screen" maximises the window. Windows has no general full-screen mode for apps.
• Echo cancellation: not done yet. While music plays loudly the wake word hears it; Alfred lowers other apps while it listens.