deepgram / deepclaw

The single skill it ships, deepclaw-voice, is little more than an instruction: copy the skill folder into ~/.openclaw/skills/ and tell your OpenClaw "I want to call you on the phone," and it walks you through creating a Deepgram account, buying a Twilio or Telnyx number, and wiring the two together. The actual work — bridging a phone call's audio to Deepgram's Voice Agent API (Flux for semantic turn-taking, Aura-2 for the reply) and proxying the conversation to a local OpenClaw gateway — lives in deepclaw's own Python server, not in the skill file.

This costs real, ongoing money on top of a coding agent you're already paying for: a phone number from Twilio (about a dollar a month, $0.085/min) or Telnyx (roughly fifty cents to two dollars a month, cheaper per minute), plus Deepgram usage once its $200 signup credit runs out. It is for someone who already has OpenClaw running locally and wants a phone line into it, not a general Deepgram or voice-agent tutorial.

READ THE FULL ANALYSIS＋

The security gaps are the author's own list, not ours. The README names four: the local chat-completions endpoint the bridge proxies to is unauthenticated, so anyone who finds your ngrok URL spends your assistant's own API credits; incoming Twilio webhooks aren't verified as actually coming from Twilio; credentials sit in a plaintext .env file; and running ngrok exposes your machine to the internet for as long as it's up. The README's own advice is to keep that tunnel URL private and only run it when needed.

A separate, upstream limitation. OpenClaw's own chat-completions endpoint buffers for about five seconds before it starts streaming a reply, so every spoken response through deepclaw carries that delay regardless of which voice provider sits behind it — the README calls this out as OpenClaw's limitation, not something a different TTS or STT choice would fix.