An assistant where you already chat
Another app to open, another chat window. What if you could simply message your assistant on Telegram or WhatsApp? That was Clawdbot’s idea: an open project connecting messaging apps to a model and tools. You may recognize its lobster mascot. A note for anyone returning to this January article: when we revised it on October 4, 2026, the project was already called OpenClaw, and the old repository redirected to the new one.
You run your own gateway on a computer or server. It connects conversations to your chosen model. “Local” describes where the gateway runs, however, not the entire communication chain. Choose a cloud model and the messages it needs go to its provider. The messaging app uses its own servers too.
What sits between a message and an answer?
The gateway receives a message, passes the task along and sends the result back to the messaging app. In the original project, this involved several parts:
Working files, configuration and history can stay on your server. Data sent to the model provider and chat platform takes a separate route. The project documentation explicitly distinguishes these two things. Do the same before connecting company information.
From one chat app to several
The original overview listed the following channels. For a new installation, check each channel’s current documentation, as connection methods and requirements can change:
We would start with one channel. Sending a message from a phone and receiving a result is a good first goal. Continuity across different messaging apps needs checking in your particular session setup; connecting two channels alone does not promise it.
What can you connect beyond chat?
Voice conversations
The project described Voice Wake and Talk Mode for voice interaction. If that is your main reason to install it, first check support for your device, microphone and chosen way of starting a conversation. Do not assume identical behavior across operating systems.
Canvas: somewhere to show a visual result
Canvas displays visual elements alongside a conversation. A use is easy to imagine: request a plan and see it as a diagram rather than another long paragraph. That is an example use, not an account of our own test.
Working in the browser
Browser tools let an agent act on websites. A message from a chat app can therefore start a task that requires a browser. Separate reading a page from submitting a form or changing data from the outset.
Skills and connections to other services
Skills are instructions describing how an agent works and uses tools. The older article suggested the connections below. Check their availability and permissions in the current catalog:
The old installation instructions: what changed?
We retain the commands from the original article below to show what installing Clawdbot looked like. This is a historical record, not instructions for a new OpenClaw installation. For that, use the current documentation linked at the end.
npm install -g clawdbot@latest
clawdbot onboard --install-daemonThe wizard at the time helped select channels and configure keys, and the article specified Node.js 22 or later. Check today’s requirements and startup method in the project documentation before installing.
This is how the model was specified in the old ~/.clawdbot/clawdbot.json file. We are keeping the historical example; do not assume the schema is compatible with the current version:
{
"agent": {
"model": "anthropic/claude-sonnet-4-5"
}
}Your own server still needs access rules
Before connecting tools, decide who can give instructions and what the agent can do. The project includes mechanisms such as:
If the requirement is to send no data to external services, assess the whole setup: model, conversation channel and connected tools. Running the gateway yourself does not meet that requirement on its own.
Does this fit your day?
The appealing idea is an ordinary one: an assistant waiting where you already send messages. You do not have to open another app to give it a small job. The question is whether that convenience is worth maintaining another tool.
Before choosing, we would answer four questions:
Where to start with your own team
Instead of connecting every system at once, outline a small trial. For example: a private channel, a few people and a task using material you can share with the model. Then work through these steps:
First, one thing you really want to delegate
We would choose a task that comes back several times a week and one messaging app. Only after checking the results would we connect more tools. That makes it easier to see whether the assistant saves work or merely moves it into configuration.
Revision of October 4, 2026: we also corrected the earlier claim that a local gateway means no data is sent to external services.
Documentation and further reading:
Have a task like this in mind for your team? Our contact details are in the footer. Start by describing one specific activity.