Concierge: WhatsApp service that knows when a human should take over

I spent two decades building search engines for companies that couldn’t afford to get it wrong: Magazine Luiza, Itaú, Vivo, Petrobras. One lesson from that stretch that nobody mentions on an AI keynote stage: the most sophisticated system in the world is useless if it doesn’t know when to get out of the way.

That’s the problem Concierge solves. It isn’t another WhatsApp bot promising to automate 100% of customer service — a promise most business owners have learned to distrust on sight. It’s an orchestration daemon that runs the conversation right up to the point a human needs to step in, and then it shuts up.

Today, January 30, 2026, Concierge goes open source. The code is at github.com/ArvorCo/concierge, ready for anyone to run, audit, or contribute to.

What Concierge actually does differently

Automated WhatsApp service isn’t a new problem. What breaks most implementations is the middle: the moment a customer stops wanting to talk to a bot, or the moment a business owner wants to jump into the conversation without fighting the system to get control back.

Concierge is built around exactly that friction point:

• Per-user isolated sessions: every conversation runs in its own context, with no state leaking between different customers — essential when the same number is handling dozens of simultaneous conversations.
• Automatic pause on human presence: the instant a human writes into the conversation, the daemon detects it and steps back. No special command, no manual flag. The human owns the thread again; the bot becomes a listener.
• Native business hours: Concierge knows the difference between answering at 2pm and responding to an anxious customer at 3am — and behaves according to the business’s own rules.
• Keyword-based escalation: when a conversation moves into sensitive territory — complaints, cancellations, language signaling frustration — Concierge escalates to a human before the problem becomes a crisis.

That’s orchestration architecture, not a canned-response script wearing an AI costume.

Built on OpenClaw, engineered for production

Concierge runs on top of OpenClaw, the multi-channel messaging protocol Arvor integrates against — we didn’t reinvent message transport, and we have no interest in pretending we did. The value we built is the orchestration layer on top: session management, pause rules, escalation logic, and the set of operational decisions that separate a portfolio project from a system that survives real traffic from real customers.

That distinction matters because the AI agent space is full of polished demos that don’t survive day one in production. Concierge already has: it’s running in Brazilian businesses today, serving real customers, every day, without a babysitter.

Why open-source it now

Because it’s the most honest way to prove it works. Any company can claim “production-ready” on a landing page. Few put the code out where anyone can check. Concierge is open for anyone who wants to self-host it, adapt it to their own use case, or just read the architecture and steal the good ideas — that’s how open source is supposed to work.

Arvor’s take

Arvor isn’t in the business of selling AI as magic. We’re in the business of building systems that do the boring work reliably — and WhatsApp, whether you like it or not, is where an enormous amount of the world’s customer conversation actually happens. A daemon that knows the right moment to pause is worth more than a model that “understands everything” and never learns when to shut up.

If your business already lives on WhatsApp customer service and you’re tired of choosing between a generic bot and an overloaded human team, take a look at Concierge or talk to us about agentic consulting. If you’d rather just poke at the code, it’s open, no signup, no pitch: github.com/ArvorCo/concierge.