Set up Shika AI.
Pick your system. The page shows the steps for the one you choose, and it starts on the one it thinks you use.
Before you start
Honest status: Linux is the system this was built and used on every day. The Mac and Windows versions pass automated tests that simulate the keyboard, microphone and windows, but they have not been run on real Macs or PCs by the author yet. If something breaks, please open an issue on GitHub with your system version and the text from the terminal.
- Kubuntu or any Ubuntu based system. KDE Plasma 6 is best, since it also gives you the native Plasma widgets.
- About 1 GB of free disk space, or about 2 GB if you also install the language model. The speech model alone is about 500 MB.
- A microphone.
- Internet for the install and the first model download. After that it works offline.
- macOS 12 or newer, Intel or Apple silicon.
- Python 3.10 or newer. Homebrew users can run
brew install python. - About 1 GB of free disk space.
- A microphone. You will be asked for three permissions below.
- Windows 10 or 11.
- Python 3.10 or newer from python.org. Tick "Add python.exe to PATH" in the installer.
- About 1 GB of free disk space.
- A microphone.
Install
- Open a terminal and run these three lines.
git clone https://github.com/iam-knr/desktop-buddy.git cd desktop-buddy bash install.sh
- The script asks for your password once to install packages with apt, add a rule so push-to-talk can read your keyboard, and add you to the input group.
- Log out and log back in one time. This is needed for the input group.
- The buddy starts at login from now on. To start it right away, open it from the application menu.
On KDE Plasma the installer also adds the Tasks, Sticky Note, Focus and Clock widgets through kpackagetool6. Right click the desktop, choose Add Widgets, and search for them.
- Open Terminal and run these three lines.
git clone https://github.com/iam-knr/desktop-buddy.git cd desktop-buddy bash install-mac.sh
- The script makes a private Python environment, installs the packages, downloads the speech model, and creates Shika in your Applications folder (inside your home folder).
- It also adds a login item so the buddy starts when you sign in.
- Open the app once from Launchpad or Spotlight.
Give it the three permissions
Open System Settings, then Privacy and Security. Add the app under each of these and switch it on.
| Permission | Why it needs it |
|---|---|
| Microphone | To hear you while you hold the key. |
| Input Monitoring | To notice when you press and release the talk key. |
| Accessibility | To type the text into other apps and to press shortcuts for you. |
Quit and open the app again after you change a permission. macOS only applies it on the next start.
- Open PowerShell or Command Prompt and run these lines. If you do not have git, download the ZIP from the GitHub page, unzip it, and double-click install-windows.bat.
git clone https://github.com/iam-knr/desktop-buddy.git cd desktop-buddy .\install-windows.bat
- The installer creates a private Python environment, installs the packages, downloads the speech model, and adds a Start Menu shortcut and a Startup shortcut.
- If Windows SmartScreen warns about the script, choose More info, then Run anyway. The code is in the repository for you to read first.
- Allow microphone access when Windows asks. You can check it later in Settings, Privacy, Microphone.
First run
- A small pet companion walks onto your desktop. That is the buddy. Its name is printed on a red belt.
- The first start downloads the speech model. It takes a few minutes and the buddy says when it is ready.
- Click the buddy to pet it. Right click for the menu, where you can pick a character, rename it and open the reminders.
- Say "call yourself Momo" with the command key held, or rename it from the menu.
Keys and voice
| You do | What happens |
|---|---|
| Hold Right Ctrl, speak, let go | Your words are typed where your cursor is. |
| Hold Right Alt (Right Option on a Mac), speak, let go | It runs the command you said. |
| Say "open firefox", "close tab", "lock screen", "volume up" | Opens the app or presses the shortcut for you. |
| Say "search" and then your words | Opens a web search. |
| Say "remind me at 5 to call Rahul" | Adds a reminder. The buddy runs to the middle of the screen and barks when it is due. |
| Tap the buddy while it barks | Marks the reminder done. Double-tap to snooze 10 minutes. |
If a keyboard does not have a Right Ctrl key, change DICTATE_KEY and COMMAND_KEY in the program file.
Shika AI and models
Two settings control the AI. They are plain lines in the program file, shika.py. Search the file for the name to find the line.
| Setting | What it controls | Default |
|---|---|---|
WHISPER_MODEL | Speech to text (faster-whisper) | small.en |
WHISPER_LANG | Speech language. None means auto detect. | en |
OLLAMA_MODEL | Text cleanup and voice edits (Ollama) | qwen2.5:1.5b |
OLLAMA_URL | Where Ollama is running | http://127.0.0.1:11434 |
Turn Shika AI on
The installer asks if you want the language model, and the default answer is yes. If you skipped it, run the installer again with the AI option. Or install Ollama yourself and run ollama pull qwen2.5:1.5b.
bash install.sh --with-ollama
bash install-mac.sh --with-ollama
.\install-windows.bat -WithOllama
To check it, open the dashboard health check. The row "Shika AI (local model)" shows the model name when it is answering.
Change the language model
- Pick a model from the Ollama library, or from the table below.
- Download it:
ollama pull qwen2.5:3b(use your model name). - Open the installed program file, find
OLLAMA_MODELand put your model name inside the quotes. - Restart Shika with the command in the Restart section.
nano ~/.local/share/shika/shika.py # in nano, press Ctrl+W and search for OLLAMA_MODEL
open -e "$HOME/Library/Application Support/ShikaApp/shika.py"
notepad "$env:LOCALAPPDATA\ShikaApp\shika.py"
| Model | Size | What to expect |
|---|---|---|
qwen2.5:1.5b | about 1 GB | The default. Fast on a CPU. Good for cleanup. |
llama3.2:1b | about 1.3 GB | Small and quick. Slightly simpler wording. |
qwen2.5:3b | about 1.9 GB | Better wording and edits. Slower. |
llama3.2:3b | about 2 GB | Another 3B option. Slower. |
qwen2.5:7b | about 4.7 GB | Best quality of these. Needs 8 GB or more of memory and is slow on a CPU. |
Sizes are approximate. Ollama shows the exact size when you pull a model. A bigger model is smarter but slower. If the model takes too long, Shika gives up after a few seconds and uses its basic cleanup, so choose the biggest model that still answers quickly on your computer.
Use Ollama on another computer
Set OLLAMA_URL to the address of another machine that runs Ollama, for example http://192.168.1.20:11434. Your text then goes to that machine on your own network. Shika talks to the Ollama API only. Other services, such as hosted chat APIs, need code changes.
Change the speech model
Find WHISPER_MODEL and change the name. The new model downloads the next time Shika starts.
| Model | Size | What to expect |
|---|---|---|
base.en | about 140 MB | Fastest. Misses more words and accents. |
small.en | about 500 MB | The default. A good balance for English. |
medium.en | about 1.5 GB | More accurate. Noticeably slower on a CPU. |
small or medium | about 500 MB or 1.5 GB | Multilingual. Use these for languages other than English. |
large-v3 | about 3 GB | Most accurate and multilingual. Slow without a graphics card. |
For another language, pick a multilingual model and set WHISPER_LANG = None to auto detect, or a language code such as "hi" for Hindi. Models without ".en" in the name are multilingual. Non-English use has had much less testing.
Keep your change when you update
The installer copies shika.py over the installed one, so an update replaces your edit. To keep it, make the same change in the cloned folder (shika/shika.py) before you run the installer. To go back, set the values to the defaults above and restart.
Widgets
Four small windows float on your desktop on every system. They remember where you put them.
- Tasks and reminders. Tick a task to finish it.
- Sticky notes. Make as many as you like. Press the bin twice to erase one.
- Focus timer. Choose 15, 25, 45 or 60 minutes.
- Clock. Switch between 12 and 24 hours.
Open them from the reminder panel buttons, or say "show tasks widget", "focus timer", "clock widget", "new sticky note" or "hide all widgets". From a terminal you can run the program with --widgets.
Restart and update
Stop it and start it again:
pkill -f "shika/shika.py"; nohup env QT_QPA_PLATFORM=xcb ~/.local/share/shika/venv/bin/python3 ~/.local/share/shika/shika.py --autostart >/dev/null 2>&1 &
To update, run git pull in the folder and run bash install.sh again.
Stop it and start it again:
pkill -f ShikaApp/shika.py; open ~/Applications/Shika.app
To update, run git pull in the folder and run bash install-mac.sh again.
In PowerShell:
Stop-Process -Name pythonw -Force Start-Process "$env:LOCALAPPDATA\ShikaApp\venv\Scripts\pythonw.exe" "$env:LOCALAPPDATA\ShikaApp\shika.py"
To update, run git pull in the folder and run install-windows.bat again.
Fixes
The key does nothing when I hold it
You probably have not logged out and in since the install. Run groups in a terminal. If "input" is missing, log out and in again.
Check Input Monitoring and Accessibility for the app, then quit and open it again.
Some games and admin windows block key listeners. Try a normal window first. Then check that no other tool is using the same key.
It hears nothing
Check the microphone in your system sound settings and say a few words to see the level move. On a Mac and on Windows, also check the microphone privacy setting for the app.
The words are wrong
The default model is small.en. A larger model is more accurate and slower. A smaller one is faster and misses more. Speak at a normal pace and keep the mic near you.
The first start is slow
It is downloading the model, about 500 MB. After that it starts in a few seconds.
I cannot see the buddy
It may be behind a full screen app, or on another screen. Say "come here" or restart it with the command above.
The fix text command does nothing
That part needs the language model. Install Ollama and a model, or run the installer again with the AI option. See Shika AI and models. The rest works without it.
Your data
Everything stays on your computer in plain files. Reminders, notes, timer state and settings are JSON files you can open and back up.
| System | Where |
|---|---|
| Linux | ~/.config/shika and ~/.local/share/summit |
| macOS | ~/Library/Application Support/Shika and /Summit |
| Windows | %APPDATA%\Shika and %LOCALAPPDATA%\Summit |
Uninstall
bash uninstall.sh
bash uninstall-mac.sh
.\uninstall-windows.bat
Your reminders and notes are left in place unless you delete the data folders above.