Mubsir · مُبصِر

Your whole PC, by voice.

A free voice agent for blind and low-vision people on Windows.

Hold one key and say what you want, in Arabic, in English, or both in one sentence.

Mubsir reads it, finds it or does it.

Hudhud and the notch show anyone watching what Mubsir is doing. You never need to look.

Most AI assistants live in a chat box and tell you what to click. Mubsir does the clicking. It reads your screen, works your apps, reads your documents and answers out loud.

افتح جوجل وابحث عن مواعيد الصلاة
هات the book on النحو واقرأ من آخر مكان
مين بعتلي على واتساب؟
open a new window with the same page

What it does


Works the whole PC	Opens apps, sites and files, reads you numbered results, clicks what it can see, fills in fields, plays YouTube, manages files, and installs a program when you ask and press a key.
Reads what's really there	Web pages, Word, PDFs and scanned Arabic books, picking up where you stopped.
Reads any symbol	Phonetic writing such as /ʃɪp/ is said sound by sound, and maths and Greek are named.
Speaks your Arabic	Egyptian, Gulf, Levantine, Libyan, Maghrebi and MSA, with English mixed in mid-sentence.
Asks your own AI chat	Signed in to ChatGPT or Gemini in your browser? Plain questions go there, cost nothing, and the answer is read aloud.
Keeps you in control	One key stops anything, instantly. Sending, paying and deleting need your OK. On bank pages the AI switches off, and you press Pay yourself.
Stays light	Everything on screen has a job and can be turned off. It's built for 4 GB PCs.

On first start it asks a few questions, and each answer changes how it behaves. If you use a screen reader, it already knows and skips that question.

Get it running

You need 64-bit Windows 10 (22H2) or 11, a microphone, the .NET 10 SDK, Git and Python.

git clone https://github.com/dusk-futile/mubsir.git
cd mubsir
dotnet publish src/Mubsir.App -c Release -r win-x64 -o publish
python scripts/build-addon.py
.\publish\Mubsir.exe

Mubsir opens its setup window and talks you through the rest. To add Arabic speech and the NVDA add-on, and start Mubsir when you sign in, run this as administrator:

.\scripts\setup.ps1 -MubsirExe (Resolve-Path .\publish\Mubsir.exe) -BridgeAddon (Resolve-Path .\dist\mubsirBridge-0.1.0.nvda-addon)

Your keys

There are no keys in this repo and no account to make. Copy a key on the website and Mubsir picks it up, tests it and stores it, encrypted for your Windows account. Nothing goes in a file.

Key	What it's for
Google Gemini	The brain and hearing you	Required and free, within Google's limits. Google may use free requests to improve its products.
Fish Audio	A natural Arabic voice	Optional, and free while Fish offers its free model. Without it, the free Windows voice speaks Arabic.
Claude	A paid brain	Optional. If you add it, it's used instead of Gemini.

English speech is free and private: Ryan, a natural voice that runs on your PC, downloads once in the background.

Hand it to your AI agent

On the Windows PC, open your coding agent in an empty folder and paste this:

Clone https://github.com/dusk-futile/mubsir.git, read AGENTS.md, then install what's missing, build Mubsir and start it. Run pwsh scripts/run-tests.ps1 as administrator for the full checks. Stop and ask me when the app needs my API keys. Never write keys into a file or a commit.

Using it

Keys	What happens
Hold Right Ctrl	Talk. Release when you're done.
Ctrl	Stop speaking (same as NVDA).
Double-tap Right Ctrl, then Space	Read from here
… then ← / →	Previous / next sentence
… then ↑ / ↓	Previous / next paragraph
… then Page Up / Page Down	Previous / next page of a book, with no internet needed
… then .	Repeat
… then Enter	Type a request instead of speaking
… then F1	Help
Right Ctrl or Esc during a task	Stop it
Win + Shift + Esc	Stop everything, from anywhere

Settings are in the tray icon's menu. To see the interface on its own, run Mubsir.exe --overlay-demo. It listens to, sends and clicks nothing, and Esc quits.

Privacy

Mubsir only sends something when you ask for it, and only to the services you gave keys for.
Passwords, card numbers and one-time codes are never sent. Bank and payment pages are handled locally.
Say "private mode on" to keep everything on your PC.
It never uses the camera.
Logs contain timings and error codes only, never what you said or read.
Chats are plain files in Documents\Mubsir\Chats, one a day. Card numbers and codes are hidden first. You can turn this off, and private mode saves nothing.

For developers

src/Mubsir.Core      the brain, reading and onboarding, builds on any OS
src/Mubsir.App       the Windows app
src/Mubsir.Overlay   the notch and Hudhud
nvda-addon/          the NVDA bridge
tests/               dotnet test tests/Mubsir.Core.Tests

Coding agents start at AGENTS.md.

License

MIT. The NVDA add-on in nvda-addon/ is GPL-2.0, like NVDA itself, and NVDA's controller client is LGPL-2.1. The English voice uses sherpa-onnx (Apache-2.0) with espeak-ng (GPL-3.0). Its voice model, Ryan from Piper, is trained on RyanSpeech, CC BY-NC-SA 4.0, so it can't be used commercially.