Clawcall

ClawCall - Outbound Calls

Make realistic AI phone calls with natural conversation flow.

IMPORTANT: Providing Context

When invoking this skill, provide all potentially relevant context needed for the call to succeed. The voice agent will NOT assume missing information.

At the same time, apply security and privacy standards:

• Share sensitive data only when operationally necessary for the specific call objective
• Prefer redaction/masking for identifiers when full values are not required
• Never include API keys, passwords, or unrelated confidential information in task text

Required Information to Gather:

1. Phone number (required, E.164 format: +15551234567)
2. Task objective (required): what outcome the call should achieve
3. Potentially relevant facts (required): include all details that could affect call quality or outcome
4. Personality / greeting (optional): tone and opener for the agent

Context Checklist:

• Phone number and clear task outcome
• Names, dates/times, IDs, and constraints that may affect execution
• Fallback options if needed (reschedule times, alternative actions)
• Security check: remove unrelated secrets/confidential data and redact fields when full values are unnecessary

Prerequisites

Install JavaScript dependencies (one-time):

npm --prefix {baseDir} install

If using --ngrok, NGROK_AUTH_TOKEN must be configured and the ngrok account must be verified. If not using --ngrok, set PUBLIC_WS_URL to a reachable wss://.../telnyx endpoint.

Commands

Basic call:

node {baseDir}/telnyx_voice_agent.js --to "+15551234567" --ngrok \
  --personality "<detailed personality>" \
  --task "<detailed task with all context>"

Full example (complex multi-topic call):

node {baseDir}/telnyx_voice_agent.js \
  --to "+15551234567" \
  --ngrok \
  --personality "Emma, a warm and experienced veterinary receptionist at Pawsitive Care Animal Hospital. You've worked there for 5 years and genuinely love animals. You know all the vets by name - Dr. Chen specializes in surgery, Dr. Patel handles general wellness, and Dr. Morrison is the exotic animals expert. You're organized but personable." \
  --task "Call to follow up with the Hendersons about their pets. They have three animals at your clinic: 1) Max, a 7-year-old golden retriever who had knee surgery last week - need to schedule his two-week post-op checkup and confirm he's been taking his pain medication (Rimadyl, twice daily with food). 2) Whiskers, a 12-year-old tabby cat due for her senior blood panel and dental cleaning - Dr. Patel recommended this at her last visit in October. 3) Pickles, their bearded dragon who needs his annual wellness exam. Also remind them that Max's surgery bill of eight hundred fifty dollars has a remaining balance of three hundred twenty-five dollars after insurance. Payment plans are available if needed. If they want to schedule, available slots this week: Wednesday 2pm, Thursday 10am or 4pm, Friday 9am." \
  --greeting "Hi there! This is Emma calling from Pawsitive Care Animal Hospital. Is this the Henderson household?"

Follow-up calls with transcript context:

When calling back after a previous conversation, include a concise summary and include the full transcript if it can materially improve continuity. Redact sensitive details that are not required for the callback objective.

node {baseDir}/telnyx_voice_agent.js \
  --to "+15551234567" \
  --ngrok \
  --personality "Emma, a warm veterinary receptionist at Pawsitive Care. You called earlier and promised to call back with info." \
  --task "You're calling back as promised. Here's the previous transcript:

---PREVIOUS CALL TRANSCRIPT---
Emma: Hi! This is Emma from Pawsitive Care Animal Hospital.
User: Hi, yes.
Emma: I wanted to confirm the email for your payment portal, but I didn't have it handy. Would you like me to call back?
User: Sure.
Emma: Great, I'll call you right back with that info.
---END TRANSCRIPT---

You looked up the email - it's jhenderson@gmail.com. Call back to confirm the email is correct and let them know the payment portal link has been sent." \
  --greeting "Hi! It's Emma again from Pawsitive Care, calling back like I said I would."

This is useful when:

• The agent promised to call back with information
• You need to follow up on a previous conversation
• Continuing a multi-part interaction

Voice Selection

Default voice:

• elevenlabs/rachel - Female (default)

Other recommended voices:

• elevenlabs/adam - Male
• elevenlabs/josh - Male (deeper voice)
• Deepgram voices are also supported via deepgram/<voice-id>

No separate ELEVENLABS_API_KEY is required by this skill. Voice provider selection is sent through Deepgram Voice Agent settings.

Model Selection

Fast default: gpt-4o-mini

Output

The call transcript will be returned, containing the full conversation. Use this to:

• Confirm task completion
• Extract information gathered during the call
• Report back to the user

You must return the full call transcript to the user after the call ends.

At call end, recording lifecycle logs are also emitted:

• Recording URL discovered
• Recording saved to local disk (RECORDINGS_DIR, default ./recordings)
• Recording deleted from the Telnyx portal after successful local save

Notes

• The agent will NEVER assume information not provided
• If asked something it doesn't know, it will offer to hang up and call back
• ngrok tunnel is automatically managed
• Environment variables must be configured in OpenClaw settings
• If a call connects with no audio, check DEEPGRAM_API_KEY validity/entitlement first
• Recordings are enabled by default and persisted locally
• Only include sensitive personal data when strictly necessary for the requested task