Voice Calls (OpenAI Realtime)

Make real phone calls through your OpenClaw agent. Ask it to book a restaurant, check store hours, schedule an appointment — it dials the number, handles the conversation, and reports back with structured results.

Uses OpenAI's Realtime API for single-model speech-to-speech with ~200-300ms response latency. No separate STT or TTS — one model does it all.

Setup

This skill requires a Twilio account and an OpenAI API key with Realtime API access.

1. Set your credentials in the plugin config (via OpenClaw settings or openclaw.json):
• twilio.accountSid — your Twilio Account SID
• twilio.authToken — your Twilio Auth Token
• fromNumber — a Twilio voice-capable phone number (E.164 format, e.g. +17075551234)
• openai.apiKey — your OpenAI API key
• publicUrl — a public HTTPS origin that routes to the plugin's server (port 3335 by default). Must not be localhost/private/internal.

1. Set up a tunnel (Cloudflare Tunnel, ngrok, Tailscale Funnel, etc.) so Twilio can reach the webhook server.

1. Verify setup:

openclaw voicecall-rt status

Usage

Just tell your agent what to call and why:

> "Call Tony's Pizza at +14155551234 and reserve a table for 4 on Friday at 7pm"

> "Call the barbershop at +14155559876 and book a haircut for Saturday morning"

> "Call +14155550000 and ask if they have the iPhone 16 Pro in stock"

The agent writes a system prompt for the voice AI, dials the number, and the voice AI handles the conversation autonomously — including navigating phone menus (DTMF), detecting voicemail, and reporting the outcome. The plugin wraps prompts with safety guardrails and blocks deceptive identity behavior.

CLI

openclaw voicecall-rt call -n +14155551234 -t "Check store hours"
openclaw voicecall-rt status
openclaw voicecall-rt active

Inbound calls

Optionally receive calls by enabling inbound.enabled and setting a policy (open or allowlist). Disabled by default.

Cost

~$0.31/min total (~$0.06 OpenAI input + ~$0.24 OpenAI output + ~$0.014 Twilio). A typical 5-minute call costs ~$1.55.

Notes

• The voice AI waits for the callee to speak before talking ("listen first") — no awkward overlap on pickup.
• Server binds to 127.0.0.1 by default. Only exposed via your tunnel.
• Max 5 concurrent calls by default (configurable via calls.maxConcurrent).
• Debug mode (debug: true) enables call recording, verbose logging, and latency metrics; recordings/transcripts may contain sensitive data.