Talk, and it types.
Hold Right Ctrl, speak, let go. The text lands at your cursor in whatever app you are using. No account and no internet needed after the first model download.
Shika AI is a small pet companion that lives on your screen. Hold a key and talk, and your words are typed into any app. It also keeps your reminders and barks when one is due.
Speech is turned into text on your own computer. Nothing is sent to a server. Also known as Desktop Buddy.
Hold Right Ctrl, speak, let go. The text lands at your cursor in whatever app you are using. No account and no internet needed after the first model download.
Hold Right Alt (Right Option on a Mac) and say what you want.
open firefoxclose tablock screenvolume upsearch DV360 line item setupbe a pandacall yourself MomoSay "remind me at 5 to call Rahul". At 5 the buddy runs to the middle of your screen and barks. Tap it to mark the reminder done, double-tap to snooze 10 minutes.
Select text in any app, hold the command key and say "fix grammar" or "make it shorter". A small model running on your own machine rewrites it and pastes it back. Ctrl+Z undoes it. This part is optional and needs Ollama.
Shika AI uses two small models. Both run locally, so your voice and your text are not sent to any company's server. You can swap either one for a model you like better.
Turns your voice into text. It uses faster-whisper with the small.en model. The installer downloads it, about 500 MB, and it runs on your CPU.
Cleans up what you said and edits selected text. It uses Ollama with qwen2.5:1.5b, about 1 GB. The installer asks if you want it, and the default answer is yes.
Use a bigger, smaller or multilingual model. Change one name in the program file and restart. The setup guide shows each step.
Change the modelum so remind me at 4 no wait 5 to call Rahul comma and send the invoice
Remind me at 5 to call Rahul, and send the invoice.
Results depend on the model you pick. This is an example, not a promise for every sentence.
Dictation, commands, reminders and widgets do not need it. Without it, Shika uses quick rule based cleanup. Voice edits like "make it shorter" and "fix grammar" do need the language model.
# change these two lines in shika.py, then restart WHISPER_MODEL = "small.en" # speech to text OLLAMA_MODEL = "qwen2.5:1.5b" # text cleanup and edits
Three pups and twelve other pet companions, each drawn in code with its own moods. Pick one below and the buddy at the bottom of this page changes with it. Every character keeps its own name.
Tasks, sticky notes, a focus timer and a clock. They float on your desktop, remember where you left them, and work the same on Linux, macOS and Windows. These are live samples.
On Linux with KDE Plasma you can also use the same widgets as native Plasma widgets. Both read and write the same files.
Linux is where it was built and tested every day. The Mac and Windows versions came later and have only been tested with simulated keyboards and windows, not on real machines yet. Please report what breaks.
Needs Microphone, Input Monitoring and Accessibility permissions once.
Mac guideOne installer. Start Menu entry and start at sign-in.
Windows guideDictate emails, captions and drafts. The cleanup step removes the "ums" and fixes the punctuation for you.
Speak notes while you read. Keep tasks and reminders on the desktop, with a focus timer next to them.
Dictate comments and messages, open apps and close tabs by voice, and read or change the code. It is one Python file.
| Shika AI | Typical cloud dictation | |
|---|---|---|
| Where your voice goes | Stays on your computer | Sent to a server |
| Works offline | Yes, after setup | Usually no |
| Account needed | No | Usually yes |
| Price | Free, MIT license | Often a subscription |
| Choose your own model | Yes | No |
| Source code | Open | Closed |
| Accuracy on hard audio | Good, depends on the model you pick | Often better on very noisy audio |
You need Python 3.10 or newer on Mac and Windows. The first run downloads a speech model of about 500 MB. The installer then asks if you want Shika AI, the language model, which is about 1 GB more. Add --no-ai to skip the question.
git clone https://github.com/iam-knr/desktop-buddy.git cd desktop-buddy bash install.sh
Log out and back in once so push-to-talk can read your keyboard.
git clone https://github.com/iam-knr/desktop-buddy.git cd desktop-buddy bash install-mac.sh
macOS asks for Microphone, Input Monitoring and Accessibility. The guide shows where.
git clone https://github.com/iam-knr/desktop-buddy.git cd desktop-buddy .\install-windows.bat
No git? Download the ZIP from GitHub and double-click install-windows.bat.
Shika AI is a free, open source desktop companion and offline voice assistant for Linux, macOS and Windows. A small pet companion lives on your screen. You hold a key, speak, and your words are typed into the app you are using. It also keeps reminders, notes and a focus timer.
Yes. Desktop Buddy is the earlier name of the same project. The pup that walks on your screen is called Shika by default, and you can rename it.
You always get the speech model, which turns your voice into text. The installer downloads it, about 500 MB. The language model that tidies your text is asked about during install, and the default answer is yes. It is about 1 GB. You can skip it with the no-ai option and add it later.
No. Speech recognition and the language model both run on your computer. The only downloads are the model files, once.
Yes, after the first install and model download. Dictation, commands, reminders and widgets all work offline.
Open the program file, find OLLAMA_MODEL for the language model or WHISPER_MODEL for the speech model, change the name, and restart. The setup guide has the exact steps, a list of good models and what each one costs in size and speed.
Keep the defaults, or use base.en for speech and a 1B language model. A bigger model is smarter but slower on a CPU. If the language model is too slow, Shika gives up after a few seconds and uses basic cleanup instead.
English by default. You can switch to a multilingual speech model and set the language, or leave it on auto. Hindi and Hinglish text is kept as you speak it. Languages other than English have had much less testing.
No. Both models run on the CPU. A graphics card is not needed.
Yes. It is open source under the MIT license. You can read the code, change it and share it.
Yes. Characters are plain entries in one list in the program file. Copy one and change the colors, ears, tail and face.
Agentic AI Consultant, Digital Marketing Trainer and Developer. India.
I run ads for brands, build AI agents that handle the boring parts of marketing, and teach both. I built Shika because I wanted a desktop buddy that could also type for me.