Skip to content
All work

2026 · Design, Electron, Win32 automation

Hitodama

An ambient companion that floats on the desktop as a dynamic island: a procedural flame, push to talk dictation, real Windows controls and any OpenAI compatible model.

A closer look

Hitodama

A desktop companion with a soul

Repository

hitodima-oss

pushed 27 Oct 2026

Commits in the last year

2

Language bytes

  • JavaScript50k
  • HTML15k
  • CSS15k
  • PowerShell1k

The problem

Desktop assistants either take over the screen or need an account. Neither is acceptable for something you keep open all day, and neither respects the fact that you already have a model running locally.

What I built

  1. 01

    The soul is a canvas shader rather than a video: it reacts to four states and carries orbiting sakura petals, so it costs almost nothing when idle.

  2. 02

    A floating dynamic island docks to the top bezel with click through hit testing and wheel gestures to expand the control centre.

  3. 03

    Five slots cover personality, voice, abilities and dictation. Abilities are real Win32 calls run through PowerShell, including dark mode, master volume, Wi-Fi state and wallpaper changes.

  4. 04

    Push to talk listens on a global shortcut and injects the transcript straight into whichever window has the cursor.

  5. 05

    The endpoint is any OpenAI compatible URL, so Ollama, LM Studio, Groq, OpenRouter and Gemini all work without special cases.

What it changed

  • No account, no subscription, and the whole thing runs offline against a local model.

  • Injection at the cursor means the transcript lands where you are typing, instead of in a chat window you then copy from.

  • The mural set ships with the app, so the first launch already looks finished.