modbender percept-voice-cmd
Voice command detection and action execution for OpenClaw agents.
What it does
Detects wake words in ambient speech and routes voice commands to your OpenClaw agent for execution. Supports email, text, reminders, search, calendar, notes, and custom commands.
When to use
- User says a wake word ("Hey Jarvis", "take notes", or custom)
- User wants hands-free control of their agent
- User asks to set up voice commands or wake words
Requirements
- percept-listen skill installed and running
- OpenClaw agent accessible via CLI
Supported actions
| Action | Example |
|---|---|
| "Hey Jarvis, email Rob saying the report is ready" | |
| Text | "Hey Jarvis, text David that I'm running late" |
| Reminder | "Hey Jarvis, remind me in 30 minutes to call the dentist" |
| Search | "Hey Jarvis, look up the weather in San Francisco" |
| Calendar | "Hey Jarvis, what's on my calendar today?" |
| Note | "Hey Jarvis, take a note that the server password changed" |
| General | "Hey Jarvis, [anything]" → forwarded to OpenClaw |
Wake word configuration
Default wake words: hey jarvis, take notes, send an email
Configure via Percept dashboard (port 8960) → Settings → Wake Words, or directly in the database.
How it works
- Percept buffers incoming transcript segments
- Wake word detected → extends buffer by 5 seconds for full command capture
- 10-second continuation window after wake word (catches follow-up without repeating wake word)
- Two-tier intent parsing: fast regex first, LLM fallback for complex commands
- Contact resolution from address book (
percept/data/contacts.json) - Command dispatched to OpenClaw CLI for execution
Speaker authorization
Only approved speakers can trigger voice commands. Configure in:
percept/data/speakers.json— map speaker IDs to names- Dashboard → Settings → Speakers → toggle approved/owner
Unapproved speakers are logged but commands are not executed.