Voxclaw

VoxClaw — Voice Output for OpenClaw

VoxClaw is a macOS app that speaks text aloud. It runs on the user's Mac as a menu bar app and listens for text over HTTP. Send text from any machine on the LAN and VoxClaw reads it using OpenAI neural voices or Apple's built-in TTS.

This is useful when your agent runs on a remote server or headless machine and you want voice output on the user's local Mac.

By default, agents should speak a short final summary after non-trivial work completes. Live progress narration should remain opt-in.

How It Works

VoxClaw runs an HTTP listener (default port 4140) on the user's Mac. To speak text, POST to the /read endpoint. The user hears the text immediately through their Mac speakers with a teleprompter overlay showing synchronized word highlighting.

Fast Handoff (Human -> Agent)

If the human shares a 🦞 VoxClaw setup pointer, use it directly. It includes:

• website pointer (https://voxclaw.com/)
• integration doc (SKILL.md)
• machine-specific Speak URL (/read)
• machine-specific Health URL (/status)

Prefer those provided URLs over guessed hostnames when both are available. Never auto-switch to .local hostnames. Use numeric LAN IP URLs unless a human explicitly provides a .local target. If health_url or speak_url are present in the pointer, do not ask for LAN IP or run discovery first; call health_url immediately, then use the provided URLs.

Reliable connect order:

1. Confirm on VoxClaw Mac: curl -sS http://localhost:4140/status
2. Confirm from agent host: curl -sS http://<lan-ip>:4140/status
3. Send all speech — including final summaries and failures — to <lan-ip>:4140/read
4. If step 1 passes but step 2 fails, treat as network/firewall issue (not app API issue).

API

Speak Text

curl -X POST http://<mac-ip>:4140/read \
  -H 'Content-Type: application/json' \
  -d '{"text": "Hello from your agent!"}'

Parameters (JSON body):

Plain text also works:

curl -X POST http://<mac-ip>:4140/read -d 'Hello from your agent!'

Response:

{"status": "reading"}

Stop Reading (Ack)

Stop reading the response you previously sent — scoped to a single agent by project_id + agent_id, so other agents keep speaking. Useful when the user sends your agent a new prompt and the old spoken response is now stale.

curl -X POST http://<mac-ip>:4140/ack \
  -H 'Content-Type: application/json' \
  -d '{"project_id": "/path/to/repo", "agent_id": "session-123"}'

The ack stops local playback and is relayed to your LAN peer speakers so it stops there too. Response: {"status":"acknowledged"}.

Check Status

curl http://<mac-ip>:4140/status

Response:

{
  "status": "ok",
  "service": "VoxClaw",
  "reading": true,
  "state": "playing",
  "word_count": 42,
  "website": "https://voxclaw.com/",
  "skill_doc": "https://github.com/malpern/VoxClaw/blob/main/SKILL.md",
  "discovery": "_voxclaw._tcp",
  "speak_url": "http://192.168.1.50:4140/read",
  "health_url": "http://192.168.1.50:4140/status"
}

States: idle, loading, playing, paused, finished.

Setup

The user installs VoxClaw on their Mac:

1. Download from GitHub Releases
2. Move to Applications, launch once to complete onboarding
3. Enable "Network Listener" in Settings (or launch with voxclaw --listen)

The listener binds to all interfaces on port 4140 by default. The port is configurable in Settings or via --port.

OpenAI API key is optional. Without a key, VoxClaw uses Apple's built-in voices. With a key, it uses OpenAI's neural voices (the user provides their own key during onboarding or in Settings).

Discovery

VoxClaw advertises itself via Bonjour as _voxclaw._tcp on the local network. Agents can discover it without knowing the IP address.

Errors

Error responses are JSON: {"error": "description"}.

CORS: The HTTP API allows requests from http://localhost only. For cross-machine access, use curl or any HTTP client directly (CORS only applies to browsers).

Examples

Speak a summary after a task completes:

curl -X POST http://192.168.1.50:4140/read \
  -H 'Content-Type: application/json' \
  -d '{"text":"Task complete. I deployed the new version and all tests passed."}'

Use a specific voice at faster speed:

curl -X POST http://192.168.1.50:4140/read \
  -H 'Content-Type: application/json' \
  -d '{"text":"Heads up, the build failed on CI.","voice":"nova","rate":1.3}'

Control speaking style with instructions:

curl -X POST http://192.168.1.50:4140/read \
  -H 'Content-Type: application/json' \
  -d '{"text": "Welcome back! Your deploy succeeded.", "instructions": "Read warmly and conversationally"}'

Check if VoxClaw is available before sending:

curl -s http://192.168.1.50:4140/status | grep -q '"status":"ok"' && \
  curl -X POST http://192.168.1.50:4140/read \
    -H 'Content-Type: application/json' \
    -d '{"text":"Ready to go."}'