clack_skill

Overview

This skill deploys and manages Clack, a WebSocket-based voice relay server that connects client audio through STT → OpenClaw agent → TTS for real-time voice conversations. It supports multiple STT/TTS providers (ElevenLabs, OpenAI, Deepgram) and on-device speech, with per-session provider selection and encrypted connections via SSL or Tailscale. Use it to add a low-latency voice interface to your OpenClaw agents while keeping control over privacy and provider costs.

How this skill works

Clack accepts PCM 16kHz mono audio over a secure WebSocket, transcribes it with the selected STT provider, forwards the text to the OpenClaw Gateway chatCompletions endpoint, then synthesizes the short LLM response via the chosen TTS provider and streams PCM back to the client. The server auto-detects available providers from configured API keys, isolates each call in a clack:<uuid> session, and enforces short response rules and limits to avoid runaway outputs. It also supports local speech mode where STT/TTS run on-device and only LLM calls traverse the server.

When to use it

• You want to add real-time voice chat capability to an OpenClaw agent.
• You need per-call control of STT and TTS providers to optimize cost or quality.
• You require encrypted voice relay without sending raw audio to third parties (use on-device mode).
• You want a pairing-based secure mobile connection workflow for iOS clients.
• You need session isolation so voice calls don't pollute persistent agent context.

Best practices

• Run with a domain and SSL (Caddy) or use Tailscale to encrypt all traffic; firewall port 9878 from public access.
• Prefer on-device STT for cost savings and privacy, combine with cloud TTS for premium voices when needed.
• Keep OpenClaw gateway chatCompletions enabled and limit server-side max_tokens to prevent long responses.
• Use the pairing system and rate limits to protect tokens; rotate RELAY_AUTH_TOKEN periodically.
• Sanitize and limit user context to avoid prompt injection; verify sanitized text returned by the server.

Example use cases

• Mobile app voice assistant: let users speak to an OpenClaw agent with per-call voice preferences.
• Privacy-first deployments: run STT/TTS on-device while routing only LLM calls through your server.
• Cost-optimized workflows: on-device transcription with premium cloud TTS to minimize API spend.
• Testing audio pipelines: use echo mode to validate round-trips without consuming LLM credits.
• Multi-voice demos: switch TTS voices per session to showcase different agent personalities.