Music Island: a music overlay and local dictation for the Windows desktop

A top edge island for media, and a shortcut for dictation

Music Island controls music from the top edge of your Windows desktop and lets you type with your voice in whatever app you are using. The project describes itself as local first with no telemetry, and the current version line is 3.0.2 for Windows 10 and 11, licensed GPL-3.0-or-later.

The island itself is a customisable strip. You drag real controls into a live layout preview, and the pieces are artwork, transport controls, progress, reactions and assistant widgets. You can put both reactions on either side, pull an edge to change the width or a corner to change the scale, and settings and play and pause always remain accessible regardless of what else you have removed. Position follows display rotation and resolution changes, and reconnecting a preferred monitor restores the island there, which is the kind of detail that matters on a multi monitor setup where the resolution changes when a laptop is undocked.

The second half of the product is dictation. You press a shortcut, speak, finish the recording, and the recognised text is inserted into the active application, with the option to keep it with its audio in History. Recognition runs on your computer. The models are downloaded for your language in Settings, and you can reuse files from Handy and manage a dictionary. There is also an optional AI processing path through a provider you configure yourself, but regular dictation stays local, provider processing is off by default, and it uses a separate shortcut so the two never collide by accident.

Portable, unsigned, and what that means for updates

The distribution model is a single executable with no installation step. You download music-island.exe, save it in a folder you want to keep and open it. Existing settings stay in the %APPDATA%\Music Island\ folder, so moving the executable around does not lose your configuration.

The readme is honest about the consequence of that choice. The executable is unsigned, so Windows may show SmartScreen. Portable updates require the release SHA-256 checksum, and you have to verify the PE before replacement. The readme makes an important distinction in a single sentence: a checksum is an integrity check, not an Authenticode publisher signature. In other words, you can prove the file has not changed in transit, but you cannot prove who published it, because there is no code signing certificate behind it. If you are the kind of user who will not run an unsigned binary, that is a hard stop, and the SECURITY.md file is linked for the details.

There is a related housekeeping point about the dictation models. The downloaded speech recognition files live under %APPDATA%\Music Island\dictation\, and deleting the EXE does not remove them. They are managed in System, Data, in the settings, which is worth knowing if you are cleaning up after moving to a different machine, because the largest files on disk are the ones left behind by the executable.

Media sources: SMTC by default, a CDP loop to Yandex as an option

Music Island connects to music in three ways, and the default is the least invasive. Windows SMTC is the default, which means the app follows the selected Windows media session, and Spotify works through its desktop app with no separate Spotify API login. That is the correct design for a Windows utility, since SMTC is the system's own media session mechanism.

The second option is a direct connection to Yandex Music. It works by connecting to the installed desktop client through a local CDP endpoint, and you connect it in Settings, Music source. The readme describes what it changes: with your consent, Music Island configures existing per user shortcuts and any existing autostart command for Direct launches, and the first connection may require restarting the client. Switching back to SMTC restores normal launches, and the client executable itself stays unchanged.

The honest limitation is stated in the same paragraph. This is an unofficial integration, and it can change after a Yandex update. That is the deal with a CDP based connection to an application that does not expose a supported API. If Yandex changes its protocol or its internal surface, the connection breaks, and nothing in this repository can prevent that.

The third option is the direct Spotify app path, and Shuffle, Repeat and reactions are available where the selected player supports them, which is a reasonable way to express that control availability is a property of the player rather than of the app.

The assistant widgets read your existing local logins

The Codex and Claude widgets are optional and off by default, and they are enabled separately from everything else. What they do is show remaining quota, either as compact rings or in a detailed view, and you can choose their size and position and whether they stay visible when the island closes.

The privacy design here is the interesting part, and it is a genuinely careful one. Codex runs the installed codex.exe app-server and relies on its existing login, without reading its auth files. Claude reads the existing local Claude Code OAuth sign-in only after you enable it, and then queries Anthropic directly without a Music Island proxy in the path. The readme states plainly that credentials are never returned to the interface or sent to Music Island servers. It also handles the failure case honestly, labelling stale or unavailable data rather than presenting it as a current limit, which is a small detail that tells you more about the implementation than a paragraph of assurances would.

The remaining features are in the same spirit of doing one thing locally. Better Voice cleans up your microphone, lets you hear the result and switches voice effects on or off, with a signal route view showing input, processing and output together. It is marked as Beta, and routing audio into another app requires VB-Cable, which is an external virtual audio device rather than something the application provides. There is also an optional native mini-player that sits beside the Windows system tray and follows the taskbar as it appears and hides, which is available on the primary horizontal taskbar, and the settings offer Windows autostart explicitly rather than by default.

Rust and Tauri own the native work, React owns the pixels

The architecture section is the clearest statement of the design. React 19 with TypeScript and Vite render the island and the settings, and Tauri 2 with Rust own the native media, windows, audio, usage connections and updates. The split is what makes the app feel native while the interface stays in a familiar component model.

The diagram shows the data flow in detail. Players and Spotify feed an SMTC provider, and the Yandex desktop client is reached through an opt in loopback CDP connection into a Direct provider. Both providers land in a Rust media controller, which drives the native taskbar player and exchanges Tauri events and commands with the app layer. The app layer sits behind a useIslandApp hook and owns configuration, media, usage, dictation and window lifecycles, feeding the island UI, the settings UI and the Better Voice UI. A Rust voice engine talks to the Better Voice UI over meters and commands, and the Handy Rust coordinator links to an embedded inference DLL over an ABI version 1 boundary, with typed IPC between the coordinator and the app.

That ABI boundary and the typed IPC are the parts worth noting, because the dictation engine is a separate process or module with a versioned interface. It is the reason the project documents the Handy adaptation journal. Dictation is adapted from Handy v0.9.7, at a specific commit, under the MIT licence, with selected fixes from 0.9.8 and loading notices from October 7, and the models themselves have their own licences. The third party notices file and the adaptation journal exist so that a future upgrade of Handy is a documented, reviewable step rather than a silent copy.

Where Music Island is the wrong tool

Three cases argue against it. The first is a non Windows machine. The documented platform is Windows 10 and 11, and the native layer is a Win32 taskbar host talking to Windows media sessions, so this is not a cross platform utility with a Windows build. The second is anyone who needs a signed binary or a package manager install. There is an unsigned portable executable, a checksum to verify, and no installer, which is a good fit for a portable utility and a poor fit for a managed fleet where software deployment expects a signed MSI with a publisher name.

The third case is a workflow that wants the assistant widgets but not the coupling. The widgets work by attaching to your existing local Codex and Claude Code sign-ins, and that is the right way to avoid asking for separate credentials, but it does mean the application sits next to your developer tool credentials. If your threat model says that no other application should be able to talk to your Claude Code session, then the widget is the thing to turn off rather than to configure.

On maintenance, the release history is tight, with v3.0.0 in September 2026, v3.0.1 in early October and v3.0.2 on 2026-10-12, and the last push on the same day. The 3.0 line is a large rewrite, given the layout editor, the taskbar player, the assistant widgets and Better Voice all arrived together, so a user upgrading from 2.x should read the CHANGELOG.md rather than assume continuity. The repository is disciplined about the things that usually get skipped in desktop apps, with a SECURITY.md, a THIRD_PARTY_NOTICES.md, a storybook setup for the component workshop, separate tsconfig files for the app, node and storybook builds, an import boundary check script wired into the lint task, and a test script for the build scripts themselves.

Editorial conclusion

Music Island suits a Windows user who wants media controls and dictation without a heavyweight assistant app, and who is willing to run an unsigned portable binary and accept that the Codex and Claude quota widgets read your existing local sign-ins. It does not suit macOS or Linux users, since the documented platform is Windows 10 and 11 only, and it does not suit anyone who needs a signed installer without extra work. Before you install it, read SECURITY.md, check the release SHA-256 checksum, and note that the last push was on 2026-10-12 with the newest release v3.0.2 from the same day.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. redheadesign/music-island on GitHub
5. Releases