GitHub - lucastononro/notify: 100% local, free, offline attention skill for Claude Code: plays a sound and speaks a short status update when a long task finishes, blocks, or needs a decision.

notify — Claude Code & Codex plugin

A 100% local, free, offline attention plugin for Claude Code and OpenAI Codex. When a long task finishes, hits a blocker, or needs a decision while you're away from the screen, the agent plays a sound — and optionally speaks a short, plain-English status update — to pull you back.

One repo ships the same skill + MCP server to both agents:

🔔 play_sound — a fast system-sound ping (hero, glass, sosumi, basso, …) for "glance at the screen".
🗣️ notify — plays a sound, then speaks a sequence of short "beats" with natural pauses, using a high-quality local voice.
🔒 Fully local: no API key, no network, no rate limits, no telemetry. Speech and sound come from OS built-ins.

The skill enforces speech-friendly messages (no tables, code, file paths, or markdown read aloud) — so what the agent says actually sounds like a person talking, while the detailed text stays in the chat.

The MCP server is published as a tiny uvx-runnable package, so both agents reference it path-free with the exact same command — no bundled-script paths, no per-tool substitution variables.

Install

Claude Code

/plugin marketplace add lucastononro/notify
/plugin install notify@notify-marketplace

Update later: /plugin marketplace update notify-marketplace

Codex (best-effort, untested — see note)

codex plugin marketplace add lucastononro/notify
codex plugin add notify@notify-marketplace

The subcommand is plugin add, not install. Then codex plugin list shows it (installed, enabled) and codex mcp list shows the notify server; codex plugin remove notify undoes it.

Verified on the Codex CLI: marketplace add → plugin add installs the plugin, bundles the skill, and registers the notify MCP server with the same path-free uvx command used by Claude Code. If you'd rather wire just the server up by hand:

codex mcp add notify -- \
  uvx --from "git+https://github.com/lucastononro/notify#subdirectory=mcp-server" notify-mcp

Verified end-to-end on a live codex install: marketplace add → plugin add installs the plugin, bundles the skill, and registers the notify MCP server, and asking Codex to ping plays the sound + speaks. Built per OpenAI's skills / plugin / MCP docs.

Prerequisites

uv on PATH — both agents launch the server with uvx, which builds + caches the package (and its one dependency, mcp[cli]) from git on first run. No manual venv.
- No uv? Install it: curl -LsSf https://astral.sh/uv/install.sh | sh (macOS/Linux) or winget install astral-sh.uv (Windows).
macOS — uses built-in say + afplay. Nothing to install.
Windows — built-in PowerShell speech + winsound. Nothing to install. (untested — see below)
Linux — needs a TTS tool (spd-say from speech-dispatcher, or espeak/espeak-ng) and a sound player (canberra-gtk-play, or paplay/ffplay + the freedesktop sound theme). (untested — see below)

Use

Ask the agent any of:

"notify me when the build finishes"
"ping me when you're done"
"let me know when the tests pass"
"play a sound"

…or invoke explicitly: /notify:notify (Claude Code) or $notify (Codex).

The agent calls the MCP tool automatically — e.g. a chime plus "All tests passed. The branch is ready for review."

Platform support

Platform	Status	Speech	Sound
macOS	✅ tested	`say` (auto-picks best neural voice — Ava Premium → … → Samantha)	`afplay` of `/System/Library/Sounds/*.aiff`
Windows	⚠️ best-effort, untested	PowerShell `System.Speech` (default SAPI voice)	`winsound` → closest `%WINDIR%\Media` file, falling back to a system beep
Linux	⚠️ best-effort, untested	`spd-say` → `espeak-ng` → `espeak`	`canberra-gtk-play` (freedesktop theme) → `paplay`/`ffplay`/`aplay` of `*.oga`

The named sounds (hero, glass, …) all resolve to a sensible per-platform equivalent. Every backend call is guarded — a missing voice/sound/player returns a descriptive string instead of crashing the server. The whole platform layer is one file: mcp-server/notify_mcp/server.py.

Repo layout

.
├── mcp-server/                         # the shared, uvx-runnable MCP server (one source of truth)
│   ├── pyproject.toml                  #   entry point: notify-mcp = notify_mcp.server:main
│   └── notify_mcp/
│       ├── __init__.py
│       └── server.py                   #   cross-platform server (notify + play_sound tools)
│
├── .claude-plugin/
│   └── marketplace.json                # Claude Code marketplace catalog
├── plugins/
│   └── notify/                         # Claude Code plugin
│       ├── .claude-plugin/plugin.json
│       ├── .mcp.json                   #   uvx --from git+…#subdirectory=mcp-server  notify-mcp
│       └── skills/notify/SKILL.md
│
├── .agents/
│   └── plugins/marketplace.json        # Codex marketplace catalog
└── codex-plugin/
    └── notify/                         # Codex plugin
        ├── .codex-plugin/plugin.json
        ├── .mcp.json                   #   same uvx command (Codex's bare server-map shape)
        └── skills/notify/
            ├── SKILL.md
            └── agents/openai.yaml      #   declares the MCP dep + allow_implicit_invocation

Both .mcp.json files run the same path-free command:

uvx --from "git+https://github.com/lucastononro/notify#subdirectory=mcp-server" notify-mcp

Develop locally

# Claude Code: load the plugin straight from the repo
claude --plugin-dir ./plugins/notify
claude plugin validate .

# Run / build the MCP server from local source (no git fetch)
uvx --from ./mcp-server notify-mcp

Cost & privacy

Cost: none. Speech and sound are produced entirely by OS built-ins — no API, no per-character billing, no rate limits.
Privacy: nothing leaves your machine. There is no notify server, no telemetry, no third party.

License

MIT