Restaurant Reservation

This skill should be used in 2 steps.

Step 1: Start the call

When the user provides reservation details, OpenClaw should:

1. Collect the required reservation information
2. Call the local backend endpoint:
   • POST http://localhost:3000/start-reservation
3. Save the returned call_id
4. Return immediately to the user

OpenClaw should not wait for the full phone call to finish in the same response.

Suggested response style:

I started the reservation call for the restaurant. Ask me again shortly and I’ll check the result.

Step 2: Check the result

When the user follows up with something like:

• “check the reservation result”
• “did the restaurant confirm?”
• “what happened with the reservation?”

OpenClaw should:

1. Reuse the most recent saved call_id from this conversation
2. Call the local backend endpoint:
   • GET http://localhost:3000/reservation-result/{call_id}
3. If the backend returns pending, tell the user the call is still in progress
4. If the backend returns completed, analyze the transcript and return the final reservation outcome

Inputs

The skill expects the following fields when starting a reservation:

• restaurant_name
• restaurant_phone
• reservation_name
• party_size
• date
• preferred_time
• fallback_times
• special_requests
• user_callback

Example input:

{
  "restaurant_name": "Dragon Village",
  "restaurant_phone": "+16145551234",
  "reservation_name": "Alex",
  "party_size": 5,
  "date": "Friday",
  "preferred_time": "7:00 PM",
  "fallback_times": ["7:30 PM", "8:00 PM", "8:30 PM"],
  "special_requests": "Indoor seating if possible",
  "user_callback": "+17404048938"
}

Backend API Flow

Start reservation

OpenClaw sends:

curl -X POST http://localhost:3000/start-reservation \
  -H "Content-Type: application/json" \
  -d '{
    "restaurant_name": "Dragon Village",
    "restaurant_phone": "+16145551234",
    "reservation_name": "Alex",
    "party_size": 5,
    "date": "Friday",
    "preferred_time": "7:00 PM",
    "fallback_times": ["7:30 PM", "8:00 PM", "8:30 PM"],
    "special_requests": "Indoor seating if possible",
    "user_callback": "+17404048938"
  }'

Example response:

{
  "ok": true,
  "status": "started",
  "call_id": "019d3081-418d-7000-afe9-0e38881c0ee4",
  "message": "Reservation call started."
}

Check reservation result

OpenClaw sends:

curl http://localhost:3000/reservation-result/YOUR_CALL_ID

If the call is still in progress, the backend returns:

{
  "ok": true,
  "status": "pending",
  "call_id": "019d3081-418d-7000-afe9-0e38881c0ee4",
  "message": "Reservation result not ready yet."
}

If the call has finished, the backend returns:

{
  "ok": true,
  "status": "completed",
  "call_id": "019d3081-418d-7000-afe9-0e38881c0ee4",
  "result": {
    "call_id": "019d3081-418d-7000-afe9-0e38881c0ee4",
    "status": "completed",
    "event_type": "end-of-call-report",
    "request": {
      "restaurant_name": "Dragon Village",
      "restaurant_phone": "+16145551234",
      "reservation_name": "Alex",
      "party_size": 5,
      "date": "Friday",
      "preferred_time": "7:00 PM",
      "fallback_times": ["7:30 PM", "8:00 PM", "8:30 PM"],
      "special_requests": "Indoor seating if possible",
      "user_callback": "+17404048938"
    },
    "transcript": "assistant: ...\nuser: ...",
    "summary": null,
    "assistant_name": "Restaurant Reservation Agent",
    "customer_number": "+17404048938",
    "vapi_number": "+16504510603",
    "received_at": "2026-03-27T19:51:17.466710+00:00"
  }
}

How OpenClaw Should Interpret the Completed Result

When the backend returns status: completed, OpenClaw should analyze:

• the original request inside result.request
• the final call transcript inside result.transcript
• the optional result.summary, if available

OpenClaw should then return structured JSON with these keys:

• reservation_status
• restaurant_name
• reservation_name
• party_size
• date
• requested_time
• confirmed_time
• special_requests
• notes
• confidence

Allowed values for reservation_status:

• confirmed
• unavailable
• callback_needed
• voicemail_left
• book_online
• unknown

Suggested interpretation instruction:

You are analyzing the final result of a restaurant reservation phone call.

You are given:
1. The original reservation request
2. The final phone call transcript
3. Optional summary metadata

Your task is to determine the final reservation outcome and return structured JSON.

Return ONLY valid JSON with exactly these keys:
- reservation_status
- restaurant_name
- reservation_name
- party_size
- date
- requested_time
- confirmed_time
- special_requests
- notes
- confidence

Allowed values for reservation_status:
- confirmed
- unavailable
- callback_needed
- voicemail_left
- book_online
- unknown

Rules:
- Use the transcript as the source of truth for the final outcome.
- Interpret noisy phone transcripts semantically rather than literally.
- Use the summary only if it helps clarify the transcript.
- Use request metadata for unchanged reservation fields.
- If the requested time was unavailable but another time was accepted and confirmed, set reservation_status to "confirmed" and set confirmed_time to the accepted time.
- If no reservation was completed, set confirmed_time to null.
- Do not invent details not supported by the transcript or request.
- confidence must be one of: high, medium, low.
- Output JSON only.

Conversation Memory Requirement

OpenClaw should save the returned call_id from the first message and reuse it in the follow-up message.

For this MVP, the follow-up message should check the most recent reservation call in the current conversation unless the user explicitly refers to a different call_id.

Main Scripts Used

server/app.py

Main backend entrypoint used by OpenClaw.

Responsible for:

• health check
• starting reservation calls
• returning reservation results
• receiving Vapi webhooks

server/vapi_client.py

Internal helper used by the backend to:

• build the runtime reservation prompt
• create outbound Vapi calls
• attach metadata to the call

server/webhook_handler.py

Internal helper used by the backend to:

• process final Vapi webhooks
• extract transcript + metadata
• store results by call_id

Local Setup

Start backend

cd skills/restaurant-reservation/server
uvicorn app:app --reload --port 3000

Start ngrok

ngrok http 3000

Configure Vapi webhook URL

Set the Vapi server/webhook URL to: