update
May 28, 2026
By Teun
Local notification plugin alerts Claude Code and Codex users
A new free plugin adds local audio and spoken alerts for Claude Code and OpenAI Codex. It runs offline, uses built-in operating system voices and sounds, and can notify users when a task finishes or needs a decision.
A new plugin for Claude Code and OpenAI Codex is designed to bring users back to the keyboard when an agent finishes work or gets stuck. According to the project description, the tool is free, runs fully offline, and uses local sounds and speech instead of any cloud service.
The plugin was built by lucastononro and combines the same notification skill with an MCP server for both agents. MCP, or Model Context Protocol, is the tool-calling layer that lets an agent talk to external tools in a standard way. In this case, the server provides two actions: play a short sound, or play a sound and speak a brief status update.
The project describes those actions as play_sound and notify. The first is meant for a quick “look at the screen” ping, while the second reads a short plain-English message with pauses and a local voice. The author says the message formatting is constrained so the speech sounds natural, while detailed output stays in the chat.
According to the README details, the plugin is fully local and does not require an API key, network access, rate limits, or telemetry. On macOS, it uses built-in system tools such as say and afplay. On Windows, it uses PowerShell speech and winsound. On Linux, it relies on local text-to-speech and audio tools such as spd-say, espeak-ng, canberra-gtk-play, paplay, ffplay, or a sound theme player.
The package is published as a tiny uvx-runnable server, which means Claude Code and Codex can reference the same command without hard-coded paths or per-tool substitutions. The author says both plugin setups point to the same path-free MCP command: uvx --from a GitHub subdirectory package, then notify-mcp.
The project includes marketplace install commands for both tools. For Claude Code, the plugin can be added through its marketplace and then installed. For Codex, the author says the correct subcommand is plugin add, not install, and that codex plugin list and codex mcp list can be used to confirm it is enabled.
The author also says the setup was verified on a live Codex install. In that test, marketplace add followed by plugin add installed the plugin, bundled the skill, and registered the notify MCP server. The same server can also be added by hand with a codex mcp add command if someone wants to wire it up directly.
The package is built against OpenAI’s skills, plugin, and MCP documentation, and the platform layer is kept in one server file. The project is released under the MIT license.
