Voxlert

LLM-generated voice notifications for Claude Code, Cursor, OpenAI Codex, and OpenClaw, spoken by game characters like the StarCraft Adjutant, Kerrigan, C&C EVA, SHODAN, and more.

Why Voxlert?

Existing notification chimes (like peon-ping) do a great job of telling you when something happened, but not what happened or which agent needs your attention. If you have several agent sessions running at once, you end up alt-tabbing through windows just to find the one waiting on you.

Voxlert makes each session speak in a distinct character voice with its own tone and vocabulary. You hear "Query efficiency restored to nominal" from the HEV Suit in one window and "Pathetic test suite for code validation processed" from SHODAN in another, and you know immediately what changed. Because phrases are generated by an LLM instead of picked from a tiny fixed set, they stay varied instead of becoming wallpaper.

Who this is for

Voxlert is for users who:

Run two or more AI coding agent sessions concurrently (Claude Code, Cursor, Codex, OpenClaw)
Get interrupted by notification chimes but can't tell which window needs attention
Want ambient audio feedback that doesn't require looking at a screen
Are comfortable installing local tooling (Node.js, optionally Python for TTS)

If you run a single agent session and it's always in focus, Voxlert adds personality but not much utility. If you run several at once and context-switch between them, it's meaningfully useful.

Quick Start

1. Install prerequisites

Minimum: Node.js 18+ and afplay (macOS built-in) or FFmpeg (Windows/Linux). That's enough to get started; TTS and SoX are optional.

Aspect	macOS	Windows	Linux
Node.js 18+	nodejs.org or `brew install node`	nodejs.org or `winget install OpenJS.NodeJS`	nodejs.org or distro package (for example `sudo apt install nodejs`)
Audio playback	Built-in (`afplay`)	FFmpeg so `ffplay` is on PATH	FFmpeg so `ffplay` is on PATH
Audio effects	SoX (optional)	SoX (optional)	SoX (optional)

See Installing FFmpeg and Installing SoX for platform-specific commands.

You will also want:

An LLM API key from OpenRouter (recommended), OpenAI, Google Gemini, or Anthropic. You can skip this and use fallback phrases only.
At least one TTS backend if you want spoken output instead of notifications only.

Backend	Best for	Requirements
Qwen3-TTS (recommended)	Apple Silicon or NVIDIA GPU	Python 3.13+, 16 GB RAM, ~8 GB disk
Chatterbox	Any platform with GPU	Python 3.10+, CUDA or MPS

The setup wizard auto-detects running TTS backends. If none are running yet, setup still completes, but you will only get text notifications and fallback phrases until you start one and rerun setup.

Can't run local TTS? Both backends require a GPU or Apple Silicon. Voxlert still works without TTS — you'll get text notifications and fallback phrases. Need help? Post in Setup help & troubleshooting.

2. Install and run setup

npx voxlert --onboard

The setup wizard configures:

LLM provider and API key
Voice pack downloads
Active voice pack
TTS backend
Platform hooks for Claude Code, Cursor, and Codex

For OpenClaw, install the separate OpenClaw plugin.

3. Start a TTS backend for spoken voice

Start Qwen3-TTS or Chatterbox, then run:

voxlert setup

This lets the wizard detect the backend and store it in config.

4. Verify

voxlert test "Hello"

You should hear a phrase and see a notification. If you do not hear speech, check that:

A TTS server is running
voxlert config shows the expected tts_backend

Visual notifications: Voxlert shows a popup with each phrase without extra install. On macOS you can use the custom overlay or system Notification Center. On Windows and Linux you get system toasts. Change it anytime with:
voxlert notification

From a git clone

Run npm install inside cli/, then use node src/cli.js or link it globally if you prefer. Config and cache live in ~/.voxlert (Windows: %USERPROFILE%\.voxlert).

Development

Run tests locally with:

npm test

For release-impacting changes, add a changeset before opening a PR:

npm run changeset

Supported Voices

The sc1-adjutant preview below uses the animated in-game portrait GIF from assets/sc1-adjutant.gif.

Pack ID	Voice	Source	Status
`sc1-adjutant`	SC1 Adjutant	StarCraft	✅ Available
`sc2-adjutant`	SC2 Adjutant	StarCraft II	✅ Available
`red-alert-eva`	EVA	Command & Conquer: Red Alert	✅ Available
`sc1-kerrigan`	SC1 Kerrigan	StarCraft	✅ Available
`sc1-kerrigan-infested`	SC1 Infested Kerrigan	StarCraft	✅ Available
`sc2-kerrigan-infested`	SC2 Infested Kerrigan	StarCraft II	✅ Available
`sc1-protoss-advisor`	Protoss Advisor	StarCraft	✅ Available
`ss1-shodan`	SHODAN	System Shock	✅ Available
`hl-hev-suit`	HEV Suit	Half-Life	✅ Available

More coming soon: Request a voice

voxlert voice

Integrations

Claude Code

Installed through voxlert setup. Claude Code hook events are processed by:

voxlert hook

Cursor

Installed through voxlert setup, or add hooks manually in ~/.cursor/hooks.json:

{
  "version": 1,
  "hooks": {
    "sessionStart": [{ "command": "voxlert cursor-hook", "timeout": 10 }],
    "sessionEnd": [{ "command": "voxlert cursor-hook", "timeout": 10 }],
    "stop": [{ "command": "voxlert cursor-hook", "timeout": 10 }],
    "postToolUseFailure": [{ "command": "voxlert cursor-hook", "timeout": 10 }],
    "preCompact": [{ "command": "voxlert cursor-hook", "timeout": 10 }]
  }
}

Cursor Hook Event	Voxlert Event	Category
`sessionStart`	SessionStart	`session.start`
`sessionEnd`	SessionEnd	`session.end`
`stop`	Stop	`task.complete`
`postToolUseFailure`	PostToolUseFailure	`task.error`
`preCompact`	PreCompact	`resource.limit`

Restart Cursor after installing or changing hooks. See Cursor integration for details.

Codex

Voxlert uses Codex's notify config so that completed agent turns call:

voxlert codex-notify

voxlert setup can install or update the notify entry in ~/.codex/config.toml. See Codex integration.

OpenClaw

OpenClaw uses a separate plugin. See OpenClaw integration for installation, config, and troubleshooting.

Common Commands

voxlert setup                  # Interactive setup wizard
voxlert voice                  # Interactive voice pack picker
voxlert pack list              # List available voice packs
voxlert pack show              # Show active pack details
voxlert pack use <pack-id>     # Switch active voice pack
voxlert config                 # Show current configuration
voxlert config set <key> <val> # Set a config value
voxlert volume                 # Show or change playback volume
voxlert notification           # Choose popup / system / off
voxlert test "<text>"          # Run the full pipeline
voxlert log                    # Stream activity log
voxlert uninstall              # Remove installed integrations
voxlert help                   # Show full help

How It Works

flowchart TD
    A1[Claude Code Hook] --> B[voxlert.sh]
    A2[OpenClaw Plugin] --> B
    A3[Cursor Hook] --> B
    A4[Codex notify] --> B
    B --> C[src/voxlert.js]
    C --> D{Event type?}
    D -- "Contextual (e.g. Stop)" --> E[LLM<br><i>generate in-character phrase</i>]
    D -- "Other events" --> F[Fallback phrases<br><i>from voice pack</i>]
    E --> G{TTS backend?}
    F --> G
    G -- Chatterbox --> G1[Chatterbox TTS<br><i>local speech synthesis</i>]
    G -- Qwen3 --> G2[Qwen3-TTS<br><i>local speech synthesis</i>]
    G1 --> H[Audio processing<br><i>echo · normalize · post-process</i>]
    G2 --> H
    H --> I[(Cache<br><i>LRU, keyed by phrase + params</i>)]
    I --> J[Playback queue<br><i>serial via file lock</i>]
    J --> K[afplay / ffplay]

A hook or notify event fires from Claude Code, Cursor, Codex, or OpenClaw.
Voxlert maps it to an event category and loads the active voice pack.
Contextual events such as task completion or tool failure can use the configured LLM to generate a short in-character phrase.
Other events use predefined fallback phrases from the pack.
The chosen phrase is synthesized by the configured TTS backend.
Audio is optionally post-processed, cached, then played through a serialized queue.

What does it cost?

The LLM step (turning events into in-character phrases) uses a small, cheap model — not Claude. Each notification costs a fraction of a cent via OpenRouter, or zero if you use a local LLM. TTS and audio run entirely on your machine at zero cost. You can also skip the LLM entirely and use only fallback phrases from the voice pack (no API key needed).

Fully local mode (no cloud at all)

Voxlert supports local LLM servers for the phrase generation step. Run voxlert setup and choose "Local LLM (Ollama / LM Studio / llama.cpp)". Any OpenAI-compatible local server works:

Server	Default URL
Ollama	`http://localhost:11434/v1`
LM Studio	`http://localhost:1234/v1`
llama.cpp server	`http://localhost:8080/v1`

Combined with local TTS (Qwen3-TTS), this gives you a completely offline setup — no API keys, no cloud, no cost.

Configuration

Run voxlert config path to find config.json. You can edit it directly or use voxlert setup and voxlert config set.

Field	Type	Default	Description
`enabled`	boolean	`true`	Master on/off switch
`llm_backend`	string	`"openrouter"`	LLM provider: `openrouter`, `openai`, `gemini`, `anthropic`, or `local`
`llm_api_key`	string \| null	`null`	API key for the chosen LLM provider
`llm_model`	string \| null	`null`	Model ID (`null` = provider default)
`openrouter_api_key`	string \| null	`null`	Legacy alias used when `llm_backend` is `openrouter` and `llm_api_key` is empty
`openrouter_model`	string \| null	`null`	Legacy alias used when `llm_model` is empty and backend is `openrouter`
`chatterbox_url`	string	`"http://localhost:8004"`	Chatterbox TTS server URL
`tts_backend`	string	`"qwen"`	TTS backend: `qwen` or `chatterbox`
`active_pack`	string	`"sc1-kerrigan-infested"`	Active voice pack ID
`volume`	number	`1.0`	Playback volume (0.0-1.0)
`categories`	object	—	Per-category enable/disable settings
`logging`	boolean	`true`	Activity log in `~/.voxlert/voxlert.log`
`error_log`	boolean	`false`	Fallback/error log in `~/.voxlert/fallback.log`

Event categories

Event categories apply across Claude Code, Cursor, Codex, and OpenClaw where the corresponding event exists.

Category	Hook Event	Description	Default
`session.start`	SessionStart	New session begins	on
`session.end`	SessionEnd	Session ends	on
`task.complete`	Stop	Agent finishes a task	on
`task.acknowledge`	UserPromptSubmit	User sends a prompt	off
`task.error`	PostToolUseFailure	A tool call fails	on
`input.required`	PermissionRequest	Agent needs user approval	on
`resource.limit`	PreCompact	Context window nearing limit	on
`notification`	Notification	General notification	on

Omitted categories default to enabled. Set any category to false to disable it:

voxlert config set categories.task.complete true
voxlert config set categories.task.acknowledge false
voxlert config set categories.session.start true

Logging

Activity logging is on by default and writes one line per event to ~/.voxlert/voxlert.log.
Error logging is off by default and records fallback situations in ~/.voxlert/fallback.log.
Debug logging for hook sources is written to ~/.voxlert/hook-debug.log.

Useful commands:

voxlert log
voxlert log on
voxlert log off
voxlert log path
voxlert log error on
voxlert log error off
voxlert log error-path

You can also manage configuration interactively with the /voxlert-config slash command in Claude Code.

Integration behavior

voxlert setup installs hooks for Claude Code, Cursor, and Codex.
Re-run setup anytime to add a platform you skipped earlier.
voxlert uninstall removes Claude Code, Cursor, and Codex integration.
OpenClaw is managed separately through its plugin.
The global enabled flag disables processing everywhere; there is no separate per-integration toggle in config.json.

Full CLI Reference

voxlert setup                  # Interactive setup wizard (LLM, voice, TTS, hooks)
voxlert hook                   # Process hook event from stdin (Claude Code)
voxlert cursor-hook            # Process hook event from stdin (Cursor)
voxlert codex-notify           # Process notify payload from argv (Codex)
voxlert config                 # Show current configuration
voxlert config show            # Show current configuration
voxlert config set <k> <v>     # Set a config value (supports categories.X dot notation)
voxlert config path            # Print config file path
voxlert log                    # Stream activity log (tail -f style)
voxlert log path               # Print activity log file path
voxlert log error-path         # Print error/fallback log file path
voxlert log on | off           # Enable or disable activity logging
voxlert log error on | off     # Enable or disable error logging
voxlert voice                  # Interactive voice pack picker
voxlert pack list              # List available voice packs
voxlert pack show              # Show active pack details
voxlert pack use <pack-id>     # Switch active voice pack
voxlert volume                 # Show current volume and prompt for new value
voxlert volume <0-100>         # Set playback volume (0 = mute, 100 = max)
voxlert notification           # Choose notification style (popup / system / off)
voxlert test "<text>"          # Run full pipeline: LLM -> TTS -> audio playback
voxlert cost                   # Show accumulated token usage and estimated cost
voxlert cost reset             # Clear the usage log
voxlert uninstall              # Remove hooks from Claude Code, Cursor, and Codex, optionally config/cache
voxlert help                   # Show help
voxlert --version              # Show version

Platform Notes

Windows: Install Node.js and FFmpeg. Ensure the npm global bin directory is on PATH so hooks can find voxlert or voxlert.cmd.
Linux: Install Node and FFmpeg so ffplay is on PATH.
macOS: Playback uses the built-in afplay; install SoX if you want optional effects and processing.

Uninstall

voxlert uninstall
npm uninstall -g @settinghead/voxlert

This removes Voxlert hooks from Claude Code, Cursor, and Codex, the voxlert-config skill, and optionally your local config and cache in ~/.voxlert.

Advanced

See Creating Voice Packs for building your own character voice packs.

Credits

Protoss Advisor voice pack inspired by openclaw/protoss-voice

Need help?

Having trouble with setup? Post in the Setup help & troubleshooting Discussion.

License

MIT - see LICENSE.

Name		Name	Last commit message	Last commit date
Latest commit History 179 Commits
.changeset		.changeset
.github		.github
assets		assets
docs		docs
openclaw-plugin		openclaw-plugin
packs		packs
qwen3-tts-server		qwen3-tts-server
skills/voxlert-config		skills/voxlert-config
src		src
test		test
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
LICENSE		LICENSE
README.md		README.md
VERSION		VERSION
config.default.json		config.default.json
package-lock.json		package-lock.json
package.json		package.json
voxlert.sh		voxlert.sh

Folders and files

Latest commit

History

Repository files navigation

Voxlert

Why Voxlert?

Who this is for

Quick Start

1. Install prerequisites

2. Install and run setup

3. Start a TTS backend for spoken voice

4. Verify

From a git clone

Development

Supported Voices

Integrations

Claude Code

Cursor

Codex

OpenClaw

Common Commands

How It Works

What does it cost?

Fully local mode (no cloud at all)

Configuration

Event categories

Logging

Integration behavior

Full CLI Reference

Platform Notes

Uninstall

Advanced

Credits

Need help?

License

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages