OpenClaw for WhatsApp: How to Set it Up and Why Most Setups Break
The promise is simple.
Take an LLM, connect it to your tools, plug it into the apps where you already talk, and suddenly you have an agent that can actually do things.
That is the OpenClaw idea.
OpenClaw is an open-source agent runtime and orchestration layer. It sits around the model and gives it the pieces it needs to behave like an agent: memory, tools, skills, plugins, channels, permissions and a Gateway that keeps the whole thing running.
That is powerful.
It also makes OpenClaw generic. It connects to Telegram, Slack, Discord, Signal, iMessage, Microsoft Teams, Google Chat, WhatsApp and a long list of other systems.
And that generality is where the trouble starts.
WhatsApp is not just another chat channel. It has its own session model, group behavior, linked-device quirks, QR login, voice notes, media handling and reconnection issues. OpenClaw can be made to work on it, but keeping it reliable is much harder than the setup guide makes it look.
So this is the honest version. What OpenClaw is, how it connects to WhatsApp, how to set it up, and why those setups stay fragile. Then the real decision: when to self-host OpenClaw, and when to use a fully hosted OpenClaw run for you, like Perisclaw.
TL;DR
|
What is OpenClaw
The LLM is the brain.
OpenClaw is the system around it. It takes messages in from a channel, gives the model memory, tools and skills, runs the agent loop, and routes the reply back out, all through a Gateway that stays alive in the background.
It runs on your own hardware. You set it up on your machine, point it at an LLM provider of your choice, and it responds across whatever platforms you have configured. The processing happens on your device, so your conversations stay with you.
It is also extensible. You bring your own model and keys, configure providers, install plugins, define skills, and decide how the agent behaves. For technically confident users, that flexibility is much of the appeal.
All of that control comes with real upkeep. You host the runtime, you maintain it, and you fix it when it breaks. (The OpenClaw docs and repo are the full reference.)
OpenClaw is not just for WhatsApp
OpenClaw is built to be generic. The same runtime connects to Telegram, Slack, Discord, Signal, WhatsApp, iMessage, Microsoft Teams, Google Chat, Matrix, LINE, WeChat, Mattermost, IRC, WebChat and SMS. The appeal is one agent that follows you across all of them.
The cost is that each channel behaves differently, and the abstractions start to leak.
Telegram has a clean bot API. Slack has workspace apps. iMessage needs macOS. WhatsApp has none of that. It runs on WhatsApp Web, QR pairing, linked-device state and Baileys.
So "OpenClaw for WhatsApp" really means one plugin and one channel inside a much larger generic framework. Worth knowing before you start.
How OpenClaw connects to WhatsApp
There are two ways to connect OpenClaw to WhatsApp, and they suit different situations.
➤ The WhatsApp Web route links OpenClaw to your account the same way WhatsApp Web does, through the plugin built on Baileys.
It works with your personal number, sets up in minutes by scanning a QR code, and is what most people use. The catch is that it depends on a live session WhatsApp can expire or flag.
➤ The official WhatsApp Business API is supported by Meta and is more stable, but it is built for business messaging, takes days or weeks to set up, needs a business number and verification, and does not behave like a personal assistant.
WhatsApp Web route | WhatsApp Business API | |
Setup time | Minutes | Days to weeks |
Personal number | Yes | No |
Supported by Meta | No | Yes |
Best for | Personal use | Business messaging |
Stability | Session-dependent | More stable |
Most people start on the WhatsApp Web route because it is quick to get running. Whether they stay on it is another question.
Here is how that route actually works. You run a login command. It shows a QR code. You scan it from your phone. The Gateway takes over the session, and messages start flowing both ways.
From there, your assistant depends on all of this at once:
The Gateway and the machine hosting it
the WhatsApp plugin and the Baileys socket
the linked-device session and QR auth state
channel config, allowlists, pairing rules and group policy
model and provider config, tools and skills
Break any one of them and the assistant goes quiet. It ignores messages, fails to send, or needs re-linking.
That list is why most setups eventually break. It is also exactly what a managed service exists to absorb.
How to set up OpenClaw for WhatsApp
This is the condensed path for the WhatsApp Web route. For the full install and every flag, use the OpenClaw docs.
You need: Node.js, a machine that stays online, an LLM provider (or local model), a spare WhatsApp number (keep it off your primary), and a way to scan a QR code.
1. Install OpenClaw and run the Gateway as a daemon. The Gateway has to stay running in the background, or the assistant cannot receive and send messages.
2. Add the WhatsApp channel.
3. Lock down access before connecting anything. You are wiring an AI agent into WhatsApp. Decide who can reach it and where it can act. Start narrow.
4. Link with QR and keep the Gateway up.
This is where headless setups hurt. On a VPS you need a clean way to see the live QR before it expires, and relayed screenshots usually go stale first.
5. Approve pairing and test. Codes expire, so approve the first request fast.
Then message the number. If the agent replies in the logs but nothing lands in WhatsApp, the problem is outbound delivery rather than the agent itself. These are the commands you will live in:
That is the setup. The fiddly parts (QR on a remote host, pairing windows, the always-on Gateway) are operational, not conceptual. They never fully go away. They become your job.
What OpenClaw can do once it is connected
With OpenClaw connected, the basics work the way you would expect. Ask a question, get an answer. Request a summary, draft a message, set a reminder, look something up.
Two things stand out beyond that.
The first is persistent memory. OpenClaw remembers facts, preferences and past context across sessions, so it builds a picture of how you work and carries it forward instead of starting fresh each time.
The second is integrations. Through skills, which are plugin-style extensions, OpenClaw can reach into the tools you already use:
Integration | What it enables |
Gmail | Read, draft and send email |
Google Calendar | Check your schedule, create events, get reminders |
GitHub | Pull requests, issues, repository updates |
Obsidian | Read and update your notes |
Browser | Web search and page retrieval |
It can also transcribe voice notes, read images and process documents, depending on the model you have connected.
All of this works well once everything is configured and the connection holds. Keeping it in that state is the part nobody factors in.
Why it breaks, and what it costs to keep alive
Getting a demo working is the easy part. Keeping it working, every day, is where it gets hard.
1. WhatsApp is not a clean bot platform
There is no official bot API for personal accounts, so the WhatsApp Web route leans on web behavior it was never meant for: QR login, linked-device state, socket reconnects.
That makes it fragile, and it also carries a real risk. Because the method is unofficial, WhatsApp can flag or restrict accounts that use it, especially when message patterns look automated.
It happens without warning, and it is your personal number on the line. That alone is a reason to use a spare number rather than your main one.
2. Sessions drop
The linked session disconnects or loops, and the assistant goes silent. You find out when you check the logs, restart the Gateway, or re-link. For something meant to catch your commitments and follow-ups, going dark at random defeats the point.
3. Hosting is yours to keep up
OpenClaw runs on your hardware, with no failover and no backup. If the machine goes down, the assistant goes down with it, and finding and fixing the cause is on you.
4. Groups and media are full of edge cases
Mentions, replies, quoted messages, group JIDs, voice notes, forwarded media, and read receipts. A generic abstraction can carry all of it, but carrying it is different from being designed for it, which is why group messages get silently dropped even when the assistant looks connected.
Media adds its own problems. The agent has to work out when an attachment is context, when it is an instruction, and when to leave it alone. Getting that right takes real product judgment.
5. Updates break working setups
OpenClaw moves fast. The Gateway, runtime, plugin, config, skills and permissions can all change under you, so an update to any layer can break your WhatsApp assistant even when it has nothing to do with WhatsApp.
Add it up, and the real cost is not the setup. It is the upkeep.
None of these failures happens once. Sessions keep dropping. Updates keep shipping. Integrations need to be reconfigured when something changes upstream.
And every time something fails, you have to work out what failed: WhatsApp, the model, the Gateway, the runtime, or your config. That cumulative time is the cost most people do not account for, and it is why a lot of OpenClaw setups quietly get abandoned a few weeks in.
Most people just want an assistant on WhatsApp, without running a runtime to get one.
Perisclaw: a fully hosted OpenClaw, built for WhatsApp
Everything above is the part you do not want to run. Perisclaw lets you skip it.
It is a completely hosted OpenClaw, built around WhatsApp and run for you. You get an always-on agent on your own number, and you never touch a Gateway, a config file, or a QR code on a VPS.
Setup takes about a minute in your browser. You sign in, link WhatsApp as a device, and say hi. It lives on your own number, in your self-chat, and you summon it anywhere else with /claw.
There is no Node to install, no daemon to run, and no dropped sessions to chase, because we keep all of that running.
Here is what you can hand off.
✅ Track what you owe, and what you are owed
Commitments on WhatsApp scroll away the second the next message lands.
Perisclaw reads the conversation and catches them in both directions: the deck you promised by Friday, the payment a vendor said they would confirm, the question you never answered. It needs no command and no format, and it brings them back in your daily digests.
✅ Keep your inbox sane
Label, summarize, mute, pin and archive in bulk, just by asking. "Archive every chat I haven't opened in 3 months." "Summarize the last 200 messages in the launch group." For a big cleanup, it proposes the plan first and waits for your OK.
✅ Run your schedule
Reminders in plain language, plus messages scheduled to send later, one-off or recurring.
Connect Google Calendar and it books meetings with a Meet link, flags clashes when someone proposes a time, briefs you before a meeting, and blocks deep-work time.
✅ Find anything
WhatsApp search matches exact words, one chat at a time. Perisclaw searches by description across every chat and connected tools, then reasons over what it finds.
Ask "what's the latest price we discussed with the vendor?" and the answer comes back with its context.
✅ Let it work on its own
Set up automated routines for recurring work that needs judgment, like "every morning, flag customer escalations from yesterday."
Hand it autonomous jobs that run over days, like "lock a 30-min slot with everyone here and send the invite."
✅ It remembers your world
Relationship memory tracks what is going on with the people you talk to, and surfaces it right before you see them.
Personal memory learns your tone, your rules, your shorthand. A personal wiki turns everything you forward to yourself into something it can actually answer from.
✅ It reaches past WhatsApp
Drive and Sheets for filing and logging. Browser automation for research and forms. Connectors for your own tools, like a CRM or a tracker. And voice, in chat or live.
✅ Or run it the other way
If you live in Claude or Cursor, Perisclaw can act as an MCP server, exposing your WhatsApp as a tool you read and act on from your AI client, with read-only or read-and-write access you control.
You stay in control throughout. Every chat has a mode: blocked, on-request, or autonomous. By default, it is read-only, and it shows you a draft before anything goes out. Your data sits in an isolated, encrypted environment and is never used to train a model.
➤ Want the full breakdown of what Perisclaw can do? Check the complete Perisclaw documentation.
OpenClaw vs Perisclaw
OpenClaw | Perisclaw | |
What it is | Open-source agent runtime you host | Fully hosted WhatsApp assistant |
Built for | Any channel, generically | WhatsApp, specifically |
Setup | Node, Gateway, QR, config files | One minute, in your browser |
WhatsApp link | WhatsApp Web / Baileys / QR you maintain | Managed and native |
Hosting | Your machine or VPS | Managed, isolated, encrypted |
Updates & uptime | Your job | Handled for you |
Monitoring | None built in | Handled |
Privacy | Local, on your hardware | Isolated, encrypted, never trained on |
Best for | Developers and self-hosters | Busy people who live in WhatsApp |
Which one is for you
➤ Choose OpenClaw if running the infrastructure is part of the appeal. You are technical, you want to self-host, inspect and modify the runtime, and run one agent across many channels. It is impressive technology, and you will own every part of keeping it alive.
➤ Choose Perisclaw if you just want the assistant. Your life already runs through WhatsApp, and you want it tracking commitments, scheduling, searching, summarizing and chasing replies, with no server to maintain. Get access at perisclaw.com.
How we built Perisclaw - Your Personal WhatsApp AI Assistant
We did not throw OpenClaw out.
We took the best of OpenClaw-style orchestration, the agent loop, the tools, memory, skills, permissions, and long-running jobs, and rebuilt the runtime around WhatsApp. Then we hosted the whole thing, so the upkeep that would normally fall on you stays with us.
OpenClaw starts generic and adds WhatsApp as one channel. Perisclaw starts with WhatsApp and runs the runtime for you.
Frequently Asked Questions
1. What is OpenClaw?
An open-source AI assistant and agent runtime that runs on your own hardware. It connects an LLM to tools, memory, skills, plugins, channels and execution environments, so the model can act like a tool-using agent that gets things done in your apps.
2. How does OpenClaw connect to WhatsApp?
Two ways. The common one is the WhatsApp Web route, which links via the plugin (built on Baileys), works with your personal number, and goes live when you scan a QR code.
The other is the official WhatsApp Business API, which Meta supports but is built for business messaging and needs verification.
3. Why does my OpenClaw WhatsApp setup keep disconnecting?
The WhatsApp Web session behaves like a browser tab, and WhatsApp expires it periodically.
When that happens, OpenClaw goes offline until you reconnect and scan a fresh QR code. It is one of the most common, recurring problems with self-hosted setups.
4. Is it safe to connect OpenClaw to my personal WhatsApp number?
The WhatsApp Web route is not officially supported by Meta, and WhatsApp has flagged and restricted accounts using unofficial protocols, especially when automated message patterns look unusual.
Understand that risk before connecting your main number. A spare number is the safer choice.
5. How much technical knowledge do I need to run OpenClaw on WhatsApp?
Enough to be comfortable on the command line, set up and maintain a server, manage dependencies, and debug config when something breaks.
It is manageable for developers and a steep barrier for anyone without that background.
6. Why does OpenClaw break on WhatsApp?
Usually linked-session drops, QR login issues, Gateway disconnects, Baileys socket problems, group policy misconfiguration, or an update changing plugin behavior.
The deeper reason is that OpenClaw is generic, and WhatsApp is a single channel inside a larger framework with a lot of product-specific edge cases.
7. Is Perisclaw built on OpenClaw?
Think of Perisclaw as a fully hosted OpenClaw, built for WhatsApp.
It takes the best of OpenClaw-style orchestration (the agent loop, tools, memory, skills, permissions, long-running jobs), rebuilds the runtime around WhatsApp, and runs the whole thing for you, so you get the power without installing, configuring or maintaining any of it.
8. Is my data safe with Perisclaw?
Every session runs in its own isolated, encrypted environment, your data is never used to train AI models, and you control exactly what the assistant can access. It is read-only by default and shows a draft before sending anything.
9. What is a fully managed alternative to self-hosted OpenClaw on WhatsApp?
Perisclaw. It runs on your own WhatsApp number, connects to Gmail and Google Calendar, and handles tasks, reminders, follow-ups and context across your conversations, with no server setup or ongoing maintenance on your end.
10. Can I use Perisclaw from Claude or Cursor?
Yes. Perisclaw runs as an MCP server, so you can read and act on your WhatsApp from Claude, Cursor and other MCP clients, with read-only or read-and-write access.