VoiceForge

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 VoiceForge?

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.

VoiceForge 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.

Quick Start

1. Install prerequisites

| 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.

2. Install and run setup

npm install -g @settinghead/voiceforge
voiceforge setup

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:

voiceforge setup

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

4. Verify

voiceforge test "Hello"

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

• A TTS server is running
• voiceforge config shows the expected tts_backend

Visual notifications: VoiceForge 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:

voiceforge 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 ~/.voiceforge (Windows: %USERPROFILE%\.voiceforge).

Development

Run tests locally with:

npm test

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

npm run changeset

Supported Voices

| | 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 | | | sc2-kerrigan | SC2 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

voiceforge voice

Integrations

Claude Code

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

voiceforge hook

Cursor

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

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

| Cursor Hook Event | VoiceForge 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

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

voiceforge codex-notify

voiceforge 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

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

How It Works

flowchart TD
    A1[Claude Code Hook] --> B[voiceforge.sh]
    A2[OpenClaw Plugin] --> B
    A3[Cursor Hook] --> B
    A4[Codex notify] --> B
    B --> C[src/voiceforge.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]

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

Configuration

Run voiceforge config path to find config.json. You can edit it directly or use voiceforge setup and voiceforge 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 | "chatterbox" | TTS backend: chatterbox or qwen | | active_pack | string | "sc2-adjutant" | 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 ~/.voiceforge/voiceforge.log | | error_log | boolean | false | Fallback/error log in ~/.voiceforge/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:

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

Logging

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

Useful commands:

voiceforge log
voiceforge log on
voiceforge log off
voiceforge log path
voiceforge log error on
voiceforge log error off
voiceforge log error-path

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

Integration behavior

• voiceforge setup installs hooks for Claude Code, Cursor, and Codex.
• Re-run setup anytime to add a platform you skipped earlier.
• voiceforge 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

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

Platform Notes

• Windows: Install Node.js and FFmpeg. Ensure the npm global bin directory is on PATH so hooks can find voiceforge or voiceforge.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

voiceforge uninstall
npm uninstall -g @settinghead/voiceforge

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

Advanced

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

Credits

• Protoss Advisor voice pack inspired by openclaw/protoss-voice

License

MIT - see LICENSE.