A local MCP notification server with neural voice, desktop, and terminal channels.
It exposes voice_notify with three independent notification channels:
desktop notification through notify-send (enabled by default)
terminal bell
local neural speech through Kokoro, with spd-say as an emergency fallback
A local MCP notification server with neural voice, desktop, and terminal channels.
2026-10-28
Estrellas
1
Herramientas
—
Disponibilidad
—
Kokoro loads lazily on the first spoken notification and remains resident. The default voice is bf_emma. Audio is sent as in-memory 24 kHz PCM to PipeWire through pw-play; notification text is never passed through a shell or sent to a remote API.
Requirements
Linux with glibc 2.38 or newer and PipeWire's pw-play
notify-send for desktop notifications
spd-say for speech fallback
Rust 1.88 or newer and a native build toolchain when building from source
approximately 314 MB for the Kokoro model and bundled voice selection
The current ONNX Runtime binary requires glibc 2.38 symbols. Ubuntu 24.04 and Debian 13 are tested; Debian 12 and Ubuntu 22.04 are not currently supported by the default build. Release binaries and packages are built on Ubuntu 24.04 and require glibc 2.39 or newer. On Debian and Ubuntu, the runtime commands are provided by pipewire-bin, libnotify-bin, and speech-dispatcher. On Fedora, they are provided by pipewire-utils, libnotify, and speech-dispatcher.
The licensed integration uses kokoro-en 0.1.4, licensed Apache-2.0. The Kokoro model weights are also Apache-2.0.
Install
Debian or Ubuntu package
Download the .deb and SHA256SUMS files for the release, verify the package, and install it:
The distribution packages install voice-notifier-mcp and voice-notifier-install-assets under /usr/bin. Runtime dependencies are installed through the package manager. speech-dispatcher is recommended rather than required because it is only the emergency fallback.
Portable archive
The release also provides a linux-x86_64.tar.gz archive containing the server, asset installer, documentation, and licenses. Extract it and copy the two executables to a directory on PATH, then run voice-notifier-install-assets.
The asset installer downloads the official model and six voices, verifies their SHA-256 hashes, and places them under ${XDG_DATA_HOME:-$HOME/.local/share}/voice-notifier-mcp. Existing verified assets are not downloaded again. Once installed, synthesis is fully local.
The installer intentionally provides a small verified selection. To add another compatible Kokoro voice, obtain its raw little-endian float32 .bin embedding from a trusted source and copy it into the asset directory:
The filename stem becomes the voice_name value and must contain only ASCII letters, digits, _, or -. PyTorch .pt voice files are not directly supported; obtain or convert the corresponding raw .bin embedding rather than merely renaming the file. Verify the source, license, and checksum before installation. Restart every MCP client after adding, replacing, or removing voices so it reloads both the tool schema and voice data.
Override the defaults with VOICE_NOTIFIER_MODEL_PATH, VOICE_NOTIFIER_VOICE_PATH, and VOICE_NOTIFIER_VOICE. VOICE_NOTIFIER_VOICE_PATH may name one .bin file or a directory containing voice files. The server lists safe .bin file stems dynamically in the voice_name schema; compatible custom embeddings require no catalog update.
package-release.sh creates a portable archive, Debian package, RPM package, and SHA-256 manifest under dist/. It requires dpkg-deb and cargo-generate-rpm 0.21.0 in addition to the Rust build tools.
Tool input
{
"title": "Ferrum complete",
"message": "Lorenzo, the codebase review is done. The report is ready.",
"desktop": true,
"bell": true,
"voice": true,
"voice_name": "af_heart",
"speech_speed": 1.0
}
Only message is required. voice_name selects any installed compatible .bin voice exposed by the tool schema and defaults to bf_emma when available. speech_speed accepts 0.5 through 2.0 and defaults to 1.0. Spoken messages are limited to 500 Unicode characters. Process arguments are passed directly without a shell.
The server negotiates MCP protocol versions 2024-11-05, 2025-06-18, and 2025-11-25. Interoperability has been verified with Ferrum, Pi 0.80.3 through its MCP bridge extension, and the official MCP Inspector 0.22.0. The tool is annotated as non-destructive, non-idempotent, and local-only.
Ferrum configuration
[[mcp.servers]]
name = "voice-notifier"
command = "/home/your-user/.cargo/bin/voice-notifier-mcp"
enabled = true
Use the absolute binary path returned by command -v voice-notifier-mcp; MCP clients do not necessarily expand ~ or environment variables in command paths. Restart Ferrum after installing a new binary or changing its configuration.
Aún no hay lista de herramientas. Los servidores remotos la reciben tras la primera comprobación de estado.
Preguntas frecuentes
Preguntas
¿Cómo instalo Voice Notifier?
Copia en esta página la configuración stdio para Cursor, Claude Desktop, Claude Code o VS Code. Usa la instalación del publicador, no un proxy de PlugMCP.
¿Voice Notifier necesita una clave de API?
Este registro no indica ninguna variable de entorno obligatoria.
¿Qué transporte usa Voice Notifier?
Voice Notifier se ejecuta localmente por stdio.
¿Cuándo se comprobó Voice Notifier por última vez?
Última comprobación: 2026-10-28. A local MCP notification server with neural voice, desktop, and terminal channels.