Imagine an assistant that quietly answers your WhatsApp while you sit with a client. This week the open-source project NanoClaw drew attention because of the story behind it: its author discovered that his previous AI assistant had been silently reading every WhatsApp group on his account — including chats full of other people’s details. So he rebuilt it around one rule: the assistant only ever sees the conversation it is working on. That story is worth more to a Swiss small business than any feature list.

What NanoClaw actually is

NanoClaw is a small, free program that you run yourself. Once set up, it connects an AI — Anthropic’s Claude — to WhatsApp, Telegram, Slack, Discord and Gmail. From inside a chat you can ask questions, draft messages, set reminders or run small searches, without opening a laptop.

Two things make it unusual. First, it is tiny: the core is roughly 500 lines of code, readable in minutes. Comparable assistants are far bigger — the popular OpenClaw is reported to run past 400,000 lines, which nobody outside the project ever fully reads. Second, every conversation runs in its own sealed container — a locked room inside your computer. The assistant never sees your raw access keys and cannot wander into the rest of your machine.

In everyday terms: instead of handing an assistant the keys to your whole office, you let it work in one room, with one drawer.

Why the backstory matters more than the features

The author, developer Gavriel Cohen, wrote the first version in a weekend — 48 hours after realising his old assistant had been reading every group on his WhatsApp the whole time. Not because the tool was malicious, but because nothing had stopped it. Client names, order details, private conversations of people who never agreed to any AI reading along: all of it lay open to a program.

For a Swiss business, that lesson lands twice as hard. Under the revised Swiss data protection law you remain responsible for client data even when a tool processes it for you. And group chats are full of other people’s information — customers, staff, family — who never chose your AI assistant. NanoClaw’s strict separation is its answer to exactly that: one chat, one sealed box, no keys to anything else.

The honest costs and limits

A free tool is not a free lunch. Before you get excited, count what it takes:

  • You install and maintain it yourself. It is not an app from a store; it runs inside Docker, a sealed workshop for software. The project itself says a first setup takes 30 to 60 minutes.
  • The AI is not included. You connect your own Claude API key, and every processed message costs a small amount on Anthropic’s bill. Light use stays cheap, but the centimes add up.
  • The sandbox protects your machine, not the chat itself. When the AI “thinks” about a message, that still happens in Anthropic’s cloud — chat content still leaves your computer while it is being processed.
  • WhatsApp has strict rules. WhatsApp officially allows automated sending only through its own Business API. Third-party bridges sit in a grey zone and can, at worst, cost you your number — a real risk on a channel your clients use.

What to take with you, even if you never install it

Most readers will not spend a weekend on Docker — and that is fine. The value of the story is the checklist behind it. Before you connect any AI tool to your client communication, ask:

  1. What exactly can it read — one conversation, or everything?
  2. Where does the content go while it is processed — your computer, or a cloud abroad?
  3. What is it allowed to do — only draft, or send on its own?
  4. Who else appears in those conversations, and did they agree?
  5. Can you switch it off cleanly, and delete what it saw?

If a WhatsApp assistant ever lands on your wishlist, start small: a test number, a dummy chat, nothing client-facing in the first month. And if you connect nothing at all — this checklist still applies to the next tool that promises to answer your messages for you.