OpenClaw on WhatsApp: A Private AI in the App You Already Use
WhatsApp is the integration with the least learning curve and the most platform friction. Telegram hands you a bot token in about a minute; WhatsApp asks you to make a decision first. Here is what that decision is, before you spend an evening on the wrong route.
Two routes, and they suit different things
Linking a number is the fast path. It behaves like WhatsApp on a second device, it suits an assistant you talk to yourself, and it needs no approval from anyone. It is tied to that number and to a session that can be logged out.
The WhatsApp Business Platform is the official route. It needs a Meta business account and a phone number dedicated to it, and in exchange it is stable, supported and intended for talking to other people. Note the number is genuinely dedicated: once registered on the platform it cannot also be used in the ordinary WhatsApp or WhatsApp Business apps.
Pick by who is going to be on the other end. An assistant for you is the first route. An assistant that answers your customers is the second, and pretending otherwise ends badly.
The 24-hour window
On the Business Platform you can reply freely for 24 hours after someone last messaged you. Outside that window, you can only send pre-approved template messages. For an assistant you message yourself this is invisible, because you are always the one starting the conversation. For anything that reaches out first, such as a reminder or a notification, it is the single constraint that shapes the whole design, and it is much cheaper to learn now than after building around it.
A phone number is a commitment
Every other integration is a token you can revoke and regenerate. WhatsApp is a number, and numbers are harder to change: they are how people reach you, they carry your history, and on the Business Platform they are locked to that use. Decide deliberately whether you are giving up a number you already use, or getting one for this.
Setting it up
- Deploy OpenClaw. Launch your instance on Flux in about 30 seconds if you have not already.
- Choose your route using the section above: linked number for a personal assistant, Business Platform if other people will message it.
- Open the WhatsApp integration in your dashboard and follow the prompts for the route you picked.
- Message it and check the reply comes from your own instance before you give the number to anyone else.
Sessions expire, and that is the usual failure
On the linked-number route your connection is a session rather than a permanent token, and sessions end. Logging out from the phone ends one, so does leaving the device offline for a long stretch, and so does WhatsApp deciding the link is stale. The symptom is an assistant that answered yesterday and is silent today with nothing having changed, which is covered alongside the other quiet failures in why an assistant stops responding.
The fix is to re-link, which takes a minute. The thing worth planning for is that it will happen again, so if you are building something other people depend on, that is another argument for the Business Platform route where the credential is a token rather than a session.
What Meta sees, and what it does not
WhatsApp is known for end-to-end encryption between people, and it is worth being precise about how that applies to a bot, because the marketing shorthand is misleading in both directions.
Messages to a business or a bot are not the same as messages to a friend. On the Business Platform, content is decrypted on Meta's infrastructure so it can be delivered to the business, which is inherent in a business receiving a message at all rather than a flaw. Self-hosting therefore does not make WhatsApp end-to-end private, and any product implying it does is overselling.
What self-hosting genuinely removes is the extra party: with most WhatsApp AI bots your conversations pass through a vendor's service in addition to the platform, accumulating there under terms you did not write. Running your own instance removes that layer, and pointing it at a local model means the content is never sent to a model provider either. That is a real and meaningful reduction in who holds your data. It is not the same claim as end-to-end encryption, and conflating the two is how people end up surprised.
Media, and what your assistant can do with it
WhatsApp is a media-first platform in a way Telegram and Discord are not, and that shapes what an assistant on it is good for. Voice notes are the obvious case: dictating something while walking and getting a written summary back is a genuinely better interaction than typing it. Photographs of documents, receipts and whiteboards are the other, because the phone is already the camera.
Two practical consequences. Media arrives as files that have to be received, held while they are processed and then stored, so an assistant used this way has a heavier storage and memory profile than one handling text; the server requirements page covers how much that matters. And a model that can read images is a prerequisite rather than a nice-to-have, which is worth checking before you build a habit around it.
Why self-hosting changes the calculation
Most WhatsApp AI bots are a third-party service sitting between you and the model, which means your conversations pass through a vendor you did not choose and cannot audit. Self-hosting removes that middle party: the assistant runs on an instance you control, and the model is your pick, including a local LLM through Ollama so message content never leaves your server.
Get started
WhatsApp is one of more than 20 OpenClaw integrations. If the platform friction above is more than you want, Telegram is a bot token and five minutes, and Discord covers the community case. The full self-host guide covers the instance itself. Ready now? Deploy from the homepage.