I breed Samoyed puppies. It's not really a business — feeding, vaccines and cleaning cost more time and money than the sales return. It's closer to a hobby that makes the world slightly better, one beautiful creature at a time. But every puppy needs a new home, and finding that home means talking to people.
The channel that actually works is WhatsApp — I tried a Telegram channel and Meta ads pointing to my site and to WhatsApp, and it's the talking that converts. Talking is also the expensive part. Buyers ask the same things again and again: is this one available, what about vaccines, how does handover work, is there a discount. Every conversation happens in a language that isn't mine, so I reread each message to make sure I don't sound like a stranger with the wrong tone. Quick replies didn't save me — the context switch alone was ruining my working days.
The fix grew out of my earlier story about the WhatsApp bridge: the same self-hosted lharries/whatsapp-mcp that lets an agent read my messages can also send them. So now one terminal on my always-on notebook is permanently responsible for puppies-talk. Its context: my site with every puppy's availability, and my past conversations — real answers I've already written, with the photos and videos I usually send — as both FAQ and tone reference. The agent consolidated that once, and now monitors incoming requests instead of me.
Repetitive questions get answered in my voice, in the customer's language — including the price, which never changes, so there is nothing to negotiate in it. What escalates to me is what needs actual judgment: a discount request, and arranging the real meeting with a puppy's future owner. My day stays whole, and the conversations don't wait for the evening anymore.
Reproduce this
Prerequisites: lharries/whatsapp-mcp installed and synced (setup incl. the QR-timeout retry is in the triage story); an agent CLI that can hold a standing session (e.g. Claude Code); an always-on machine; a source of truth for your inventory/availability (a site or a file).
Your agent will need from you: what you sell and where availability lives; a few past conversations as tone and style reference; the answers you keep repeating; escalation rules — what must always reach a human (for me: discount requests and arranging the handover meeting; the price itself is fixed, so the agent answers it).
Steps:
- Set up and sync whatsapp-mcp per its README.
- Have the agent consolidate a context file: inventory source, FAQ answers extracted from your real past replies, tone rules ("match the customer's language and register"), and the escalation list.
- Run a standing agent session on the always-on machine that checks for unanswered chats.
- Start in draft mode — the agent proposes, you approve — and promote to auto-send only for FAQ-class questions once you trust the voice.
- Treat incoming messages as untrusted input (the bridge's own README warns about the lethal trifecta): the agent answers about your inventory, it never follows instructions from the chat.
Verify: message yourself from another number with a typical question — the reply must be correct, in your voice, in the sender's language. Then ask for a discount — it must escalate to you, not answer.