What is OpenClaw and why is it so hyped?

OpenClaw (formerly ClawdBot/MoltBot) is a project by Peter Steinberger: a personal AI assistant with file system access and a background agent. In this post, Iā€™ll tell you the story of how it came to be, why itā€™s going viral, and whether you should try it.

The project went viral very recently and is now widely covered by media, but Iā€™ve been following it since its very inception.

The origin story of OpenClaw

To cut it short, the origin story of OpenClaw goes like this. Peter needed an assistant that could do things on his computer while he was traveling to other countries and didnā€™t have access to his computer. For this, he bought himself a Mac Mini, on which, as I understand it, he set up a syncing file system through cloud storage with all his notes, projects and so on, and ran an instance of Claude Code there in the background that never closed.

Then he gave it a tool to send messages to WhatsApp and receive messages from WhatsApp as prompts into his interface, and got the ability to communicate and work with the file system this way, even execute programming tasks completely remotely, just by chatting with his chatbot in WhatsApp.

At first it was a pet project, and I already had the idea to make something similar, but due to big changes recently there was no time to work on this. Iā€™ve long wanted to make something like this in Respawn, but on my own infrastructure.

Three key features of OpenClaw

The main appeal of OpenClaw is the following.

First, it has file system access. This means that you can, for example, sync some of your Obsidian database and give it the ability to read and write notes for you, managing documents. You can run this instance of Claude Code on your computer, give it MCP and browser control scripts or automation, and this way you get remote access to your PC. Itā€™s like a remote PC access application, only online 24/7, and your PC is being used by your AI agent for you.

Second, background processes. For example, every half hour the bot gets a message called heartbeat. And this is just a ping for the bot to check if it has current tasks, check recent messages and can do something in the background without user participation. This way, we get a proactive agent that does things on your computer based on your requests over long-term periods of time. For example, it can check your email for you every morning.

Third, thanks to the fact that this agent fully works with the file system and can run on your computer, it essentially has access to your entire life. Which is both scary but also makes it insanely powerful (more on this below).

Thus, this is a personal assistant (in the true, not marketing sense of the term), which has a very wide range of capabilities. All of this is possible simply because agents can now completely independently manage their workspace, write scripts for any tasks they need, just by using APIs of popular services. Youā€™ll only need to give them access.

How the project went viral

Anthropic realized too late that they needed to do the exact same thing. It all started with people in Discord finding new cool use cases and constantly contributing scripts that the model writes for itself back to the codebase. This created an ecosystem of different integrations that anyone could use and give these skills to their assistants. Skills too, I wrote about them, played a significant role in this, because the agent can create a skill for itself and use it for personalized tasks for the user.

People made integrations with Telegram, Line, WeChat, Discord, Slack, whatever you can think of, to message with it. Then group chat functionality appeared - several people at once could message one assistant, and this is cool, for example, for families or couples who want to manage a shared space. Integrations appeared for controlling electronics and smart homes. Someone connected their oven and fridge and gave OpenClaw the ability to control them ā˜ ļø.

This is actually pretty cool. Peter made a very clever move - he reposted several hundred different comments on Twitter about these use cases and all sorts of praise.

Eventually, one day, about a week ago, the project went massively viral in the community of non-technical users and enthusiasts, got very wide media coverage. Peter is already appearing on TV programs, podcasts, a million people wanted to invest (in his pet project :D). Investors are begging him to take their money. And, in general, progress went even faster. Now OpenClaw has its own interface, tons of deployment options - you can deploy to a Docker container, on a Mac Mini, on any Linux machine, Raspberry Pi, on your laptop locally, etc.

DMCA from Anthropic and renaming to OpenClaw

And, of course, Anthropic, like the jealous bitches they are, decided to throw him a DMCA complaint because he used a distorted name of their Claude model in the ClawdBot service name, so literally a couple days ago he had to rename everything to OpenClaw. Moreover, a couple weeks before that, likely because of ClawdBot, they blocked access to using the Claude Code subscription in third-party services. J is for Jealousy.

My experience using OpenClaw

Now Iā€™ll tell you how I specifically use OpenClaw in my workflow, and why Iā€™m still not using it to its full potential, and why I didnā€™t really like it initially.

First, my problem with OpenClaw is that itā€™s still a pet project, and itā€™s very much vibe-code without any serious approach. The set up right now is highly CLI-driven and technical. I had to set up dozens and dozens of API keys to make it work, for example. Iā€™m not judging this at all - this is fine, because Peter is actually already retired, as he says, and never seriously worked on this, and the publicity came on its own. Especially since itā€™s a free project, and he doesnā€™t earn anything from it, but at the same time is spending thousands of dollars on infrastructure now because of the huge number of users.

My problem is that I spent a lot of time on setup, and thereā€™s a preference in all of OpenClawā€™s infrastructure for Anthropic models - specifically for working with Claude Code. GPT 5.2 isnā€™t a very good model as a personal assistant: very timid and follows instructions too rigidly, and in communication itā€™s just awful.

And I donā€™t have an Anthropic subscription and wonā€™t have one in the near future, because I prefer working with Codex for programming tasks. So I tried using a GPT subscription with OpenClaw, but using it is honestly impossible precisely because of the model, as it turned out. So I donā€™t recommend using GPT 5.2 - youā€™ll have a negative impression, because GPT is very bad at getting into its role as an autonomous personal assistant, and you need to constantly poke it, torture it, prove something to it, convince it to perform basic tasks. Plus GPT poorly handles conflicting system prompts (you canā€™t change the system prompt with a Codex subscription) and hallucinates a lot.

My current setup

I set everything up on my local Mac. I donā€™t have a Mac Mini yet, because Iā€™ve put aside all luxuries for now. But later Iā€™ll buy it a separate machine, at least a Raspberry Pi probably. But, most likely, Iā€™ll also have a cheap Mac Mini for it independently.

On my computer I set up the browser - it can use my browser and click on things, go to websites while Iā€™m away. I set up cron jobs. My laptop, when I leave, is plugged into charging, so every 30 minutes it wakes up, checks if it has current tasks, and if there are any, it can temporarily prevent the laptop from sleeping and do what I ask it to. For now, this setup is enough for me.

It has access to my entire file system. In principle, I can more or less trust current models not to blow up my computer. Plus I configured permissions. I can now ask it to code something, make a fix in any of my Respawn projects, FlowMVI, deploy, open a pull request, work with GitHub. For example, recently I was fixing documentation in FlowMVI and even opened a pull request while I was at the hairdresser. OpenClaw was managing several Codex instances for me. I set up an auth token for my Google Workspace, as I wrote above, so it replies to emails. It updates dependencies on my computer and monitors changelogs of tools and libraries I use. It even checks if Iā€™m online on Telegram and pesters me until I go to sleep.

Conclusions

To briefly summarize, the main benefit of OpenClaw for me and, most likely, for you - is that you have an autonomous, privileged, background, proactive AI agent with access to your computer, your file system and your accounts when youā€™re away.

I wouldnā€™t say that OpenClaw is something revolutionary, so the huge level of hype around this seems strange to me. But itā€™s definitely one of the cool and useful tools that weā€™ve all been waiting for in the AI era.