How to use Hermes Agent to build a WhatsApp reply bot
The author reports a working bot replying as them to two allowlisted contacts, approximately $0.005 runtime cost per reply and both Telegram and WhatsApp device-unlink kill switches.
news.ycombinator.com
Prompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Build a personal WhatsApp reply bot using Hermes Agent, Baileys and [message archive]. Allow replies only to [informed test contacts]. Retrieve eight recent replies to each contact through [vector store] before Claude drafts a response. Add [Telegram kill command] to clear the allowlist and restart the gateway.
Reach people exploring AI agents
See placements →What happened
The author reports a working bot replying as them to two allowlisted contacts, approximately $0.005 runtime cost per reply and both Telegram and WhatsApp device-unlink kill switches.
Reported by dandinu on Hacker NewsHow the agent did it
- Built a WhatsApp bot connected to the author’s primary account.
- Restricted replies to the author and an informed friend.
- Used eight recent past replies per contact to ground generated replies.
Where a person stepped in
The friend knew about the experiment. The author debugged the bridge and provided kill switches; no outgoing-message review is described.
A person set the goal and reviewed the result.
The author built a personal WhatsApp reply bot as a weekend project. It replies in their style to two allowlisted contacts: themselves and one friend who knew about the experiment. The bot connects to the author's primary WhatsApp account rather than a dedicated number.
The implementation combines Hermes Agent, Baileys, Chroma and nomic-embed-text-v2-moe, with Claude Sonnet 4.6 accessed through Azure AI Foundry. The author says the retrieval approach produced their best results: before generating each response, the bot pulls eight of the author's real past replies to that specific contact, filtered by recency. Claude then generates the reply using that retrieved material.
The author reports about two hours of work plus another hour debugging a WhatsApp multi-device LID issue. They also describe two ways to stop the bot. A Telegram command empties the allowlist and restarts the gateway in about ten seconds. Unlinking the device from WhatsApp on the phone severs the bridge session entirely in about five seconds.
The reported runtime cost is approximately $0.005 per reply. The author explicitly accepted the account ban risk associated with testing this bridge on their primary account; the post does not describe a separate review stage for outgoing replies.
Who it's for: Developers experimenting with personal messaging automation using a tightly limited group of informed contacts.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · dandinu
Frequently asked questions
What past messages does the bot retrieve?
For every reply, it pulls eight real past replies from the author to that specific contact, filtered by recency, before Claude generates the response.
How can the author stop the WhatsApp bot?
A Telegram command empties the allowlist and restarts the gateway in about ten seconds. Unlinking the device on the phone severs the bridge session in about five seconds.
More details
Remaining steps
- Debugged a WhatsApp multi-device LID issue.
- Provided a Telegram command to clear the allowlist and restart the gateway.
About this account
The bot retrieves contact-specific past replies and generates responses with Claude for two allowlisted contacts.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=48602964. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: dandinu on Hacker News
Source published 19 Jun 2026
Source checked 14 Oct 2026
Tools and topics
Hermes Agent · Baileys · Chroma · nomic-embed-text-v2-moe · Claude Sonnet 4.6 · Azure AI Foundry · Telegram
WhatsApp · Personal bot · Retrieval
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.