S

SONITURA


Hold a key and ask. It looks at your screen as you speak, answers aloud, types where your cursor waits, and points at the spot that matters.

The mind is the ChatGPT plan you already pay for — nothing more to buy

How it works

  1. Talk — Hold fn + Control and ask about whatever is on screen. It answers out loud, briefly. It listens only while the keys are held — there is no wake word.

    a spoken answer in about six seconds
    it marks the spot on screen

  2. Write — Add ⌘ and your rambling comes back written for the app you're in — a reply in Mail, a command in Terminal, a prompt in the editor — typed where your cursor stands.

    right ⌘ alone types you verbatim,
    in 0.4 seconds

  3. Act — Say “agent, …” and it clicks and types on your screen, one visible step at a time. It won't press Send on its own, and Esc stops it.

    say “code, …” for a coding agent
    in its own workspace

  4. Remember — It keeps a private profile of you, built from your own transcripts — local, plain Markdown, yours to edit or delete.

    ~/.sonitura — never uploaded

The demo

Press and hold, as you would the keys. Nothing is recorded.

Press & hold
Heard
Fig. II — the question, as recorded

Measured on an Apple-silicon MacBook: 0.4s raw dictation, about 6s for a spoken answer, about 5s per agent step.

Memory

Sonitura keeps a private profile of you, built at idle hours from the transcripts your other AI tools leave behind — the ones deleted after thirty days if nobody saves them. It learns your projects, your preferences, the way you like things done.

All of it is plain Markdown in ~/.sonitura. Open it. Read every line. Delete what you dislike. None of it leaves this machine.

profile.md example
# what it knows - prefers short words - three projects in flight — burnban, sonitura, vote.direct - reviews end-to-end; findings re-verified - open loop: the survey-builder refactor
made from 235 conversations
edit or delete anytime

Details

  • Hotkey fn + Control, held — ⌃ ⌥ works too
  • Dictation add ⌘ — rewritten for the app you're in
  • Raw dictation right ⌘ alone — typed in 0.4 seconds
  • Speech-to-text Whisper, runs on this machine — audio never leaves
  • Voice the Mac’s built-in
  • Screenshots one, at the moment you ask — never before, never after
  • Model your ChatGPT plan, via the official Codex CLI §
  • Memory built from your own transcripts; in ~/.sonitura — edit or delete
  • Safety never presses Send on its own; Esc stops everything
  • License MIT — free, open source, no account

macOS asks for Microphone, Screen Recording, Accessibility, and Input Monitoring — once each.

§ A Claude plan works too, experimentally — official CLIs either way, under your own login.

Download

Terminal

One command.

Installs to ~/.sonitura/app and gives you a sonitura command. sonitura setup takes two guided minutes.

curl -fsSL https://raw.githubusercontent.com/SYFT8/sonitura/main/install.sh | bash

Re-run to update.

Mac app

The menu-bar app.

For Apple silicon, macOS 13 or later. Drag to Applications. Not notarized yet, so open the first launch via System Settings → Privacy & Security → “Open Anyway”; macOS asks its permissions on first use.

Download

Intel Mac? Use the terminal install.

The source

Read it, fork it, build it.

Free and open source, MIT-licensed — the assistant, the screen agent, the memory, and this page are open Python and HTML you can read in an afternoon.

github.com/SYFT8/sonitura

Free  ·  MIT  ·  no account  ·  no telemetry