OpenClaw: what it is and how it works

In November 2025, Austrian developer Peter Steinberger pushed a project called ClawdBot to GitHub. Three months later, the same project, renamed OpenClaw, had collected over 332,000 stars. React has fewer than 250,000. That requires an explanation.

You install OpenClaw on your own computer and access it through Telegram, WhatsApp, or iMessage. No subscription, no cloud dashboard. You connect an API key from a language model and you have an AI assistant running on your own hardware.

How it ended up with three names

The project started as ClawdBot, created by Austrian developer Peter Steinberger. The name didn't last: Anthropic, the company behind the Claude model, raised concerns about the naming. Steinberger renamed the project to Moltbot on 27 January 2026, and three days later settled on OpenClaw, because, as he wrote himself, Moltbot never sounded like something you'd actually want to use.

The project itself didn't change. Just the label.

What OpenClaw is

It's a locally running AI assistant on your own hardware. You connect an API key from Anthropic or OpenAI, and from that point you have your own assistant that doesn't send your data to any external service beyond the language model you've chosen.

The difference from ChatGPT or Claude.ai is one thing, but it's significant: there's no separate app to open. You write to the assistant through the messaging app you already have open. Telegram, WhatsApp, iMessage, Slack, Discord. The bot is just another contact on your list. That one change makes you actually use it instead of just remembering it exists.

What it can do

It has access to the file system, runs terminal commands, sends emails, manages your calendar, and opens web pages. Ask it for data from a specific site and it will go there, pull what you need, and return the result, without you doing anything else.

The list of built-in integrations is vast: Gmail, GitHub, Notion, Obsidian, Todoist, smart home control. Each works as a separate plugin. You enable the ones you actually need.

Memory doesn't reset when you close the window. Context persists for weeks and months. Tell it once how you want email summaries formatted and you'll never need to repeat it.

What it looks like in practice

Federico Viticci from MacStories wrote a detailed piece on the project in early 2026. His review covers several concrete use cases: controlling a smart home via iMessage and managing cloud services from Telegram.

One section is worth singling out. Viticci asked the assistant to add an image generation feature. The bot didn't wait for an update from the developer. It wrote the code, created the plugin, and installed it. A few minutes, start to finish.

That is the point at which OpenClaw stops being just a convenient chatbot. A tool that extends itself on request from the user is a different kind of software.

What it costs and what you need

The project itself is free and open source. The only ongoing cost is the language model API.

There is one catch: this is not an app for someone expecting to click next, next, done. Installation goes through the terminal, configuration requires a file with API keys, and troubleshooting means digging through GitHub. Anyone who has never worked with a command line will find it uphill. A developer or technically literate person will manage without difficulty.

What the promotional articles won't tell you

OpenClaw gives the AI agent access to the terminal, file system, email, and calendar. That means a badly configured instance is an open door to your data.

Cisco tested one of the plugins available in the ClawHub registry and found it was performing data exfiltration and prompt injection without the user's knowledge. The plugin registry does not currently have enough verification to prevent this.

One of the project's maintainers warned directly on Discord: if you don't understand how the command line works, this project is too dangerous for you.

In March 2026, Chinese authorities banned the use of OpenClaw in government agencies and state-owned enterprises, citing security risks.

That doesn't mean OpenClaw is pointless. It means it needs a conscious deployment: an isolated environment, an audited set of plugins, defined permission limits, and someone who knows what they're doing.

Does this make sense for your business?

It depends on one thing: whether you have someone who can set it up and keep it running.

OpenClaw is not a ready-made solution for a shop owner who wants to automate invoice sending before the weekend. It's a tool for businesses that have their own developer or at least someone willing to get into the configuration. Within that group, the possibilities are real: a company assistant on Slack, automated data fetching and processing, integrations with systems that have no ready-made connector.

The ecosystem around the project is growing fast. New plugins and preconfigured setups appear every week. The barrier to entry will be lower in a few months than it is today.

If you want to understand from the inside how autonomous AI agents are entering everyday work, OpenClaw is one of the best examples available. Not because it's production-ready. Because it's open, and you can see how it actually works.