An open-source assistant with real access to your files, persistent memory across weeks, and no company cloud in between. Including the security problems nobody puts in the thumbnail.
Your AI assistant lives in your computer, answers you in WhatsApp, and remembers the conversation you had three weeks ago. Not in a company's cloud, not behind a paywall, not locked in their app. That is the promise of OpenClaw, and it went from launch to over 100,000 GitHub stars in two months.
The problem it addresses is real. ChatGPT, Claude, Gemini and Perplexity are powerful, but they are not in your workflow. You open a browser, type a question, wait, copy the answer, paste it somewhere else, repeat. Every conversation is isolated. And they do not live where you actually communicate.
OpenClaw runs on your own infrastructure: your computer, a Raspberry Pi, a Mac Mini or a VPS. It reaches you in WhatsApp, Telegram, Discord, iMessage, Signal or Slack. You text it like a colleague and it answers in the same thread. Because it runs locally, it has real access to your tools. In the demo I ask it to read a file from my actual file system and it does, then reasons about the contents.
Technically it is a gateway architecture: one process, multiple channels. It routes your message to whichever model you choose and connects to over 100 services through the Model Context Protocol. Memory is persistent and context spans weeks. Because it is MIT licensed, you can extend it, audit it and change how it behaves.
And now the part that belongs in every honest video about this. Security researchers from Cisco, 1Password and others raised serious concerns about prompt injection: if someone can manipulate the input, they can make the assistant run arbitrary commands. Researchers also found publicly reachable deployments with no authentication at all, exposing API keys, chat logs and system access. That happened because people followed a thread on X without understanding what they were setting up.
My honest position: run it on a separate device, sandbox it, limit its permissions and audit the code. Or do not run it and just watch the ecosystem evolve. Both are defensible. Setting it up blind on a VPS is not.
| Aspect | Detail | What it means for you |
|---|---|---|
| Licence | MIT, open source | No subscription. You pay for hosting and model API usage, nothing else. |
| Runs on | Your computer, Raspberry Pi, Mac Mini, VPS | Your data stays where you put it. No cloud logging, no tracking. |
| Reaches you via | WhatsApp, Telegram, Discord, iMessage, Signal, Slack | No extra app, no extra tab. It meets you where you already write. |
| Architecture | Gateway: one process, multiple channels | Model-agnostic. You decide which model and which APIs. |
| Integrations | 100+ services via Model Context Protocol | Real access to tools and files, not just chat. |
| Memory | Persistent, context spans weeks | You do not repeat yourself every session. |
| Known risks | Prompt injection, exposed instances, credential storage | Separate device, sandbox, restricted permissions. Not a beginner setup. |
OpenClaw is an open-source, self-hosted AI assistant that runs entirely on your own infrastructure: your computer, a Raspberry Pi, a Mac Mini or a VPS. Instead of forcing you into another interface it reaches you where you already are, in WhatsApp, Telegram, Discord, iMessage, Signal or Slack. It was previously known as Moldbot and Claudebot.
The software itself is open source under an MIT licence, so there is no subscription. What you do pay for is the infrastructure you run it on and the API costs of whichever model you connect, for example Anthropic or OpenAI. Running it on a small VPS plus API usage is the realistic cost.
ChatGPT, Claude, Gemini and Perplexity are powerful but they are not integrated into your workflow. You open a browser, ask, copy the answer, paste it somewhere else. Every conversation is isolated. OpenClaw runs on your hardware, has persistent memory that spans weeks, has real access to your files and tools, and lives in the messenger you already use.
Not automatically. Security researchers from Cisco, 1Password and others raised serious concerns about prompt injection: if an attacker can manipulate the input, they can make the assistant execute arbitrary commands. Researchers also found publicly accessible deployments with no authentication, exposing API keys, chat logs and system access. Run it on a separate device, sandbox it, limit its permissions and audit the code, or do not run it at all.
In my own setup it prepares research summaries on topics I follow and drafts scripts, which is part of automating my content pipeline. Other people use it for email triage, home automation, research and code generation. Because it reaches over 100 services through the Model Context Protocol, the practical limit is what you connect it to.
Peter Steinberger, an Austrian developer who sold his company and got bored. It launched as Claudebot in November 2025, hit 9,000 GitHub stars within 24 hours, 60,000 within the first month and passed 100,000 in two months, making it one of the fastest growing GitHub repositories ever.
Summary of all twelve chapters. Timestamps jump straight to the matching point in the video. This video has no manual subtitle track, so there is no verbatim transcript here.
The premise: an assistant that answers in WhatsApp, remembers a conversation from three weeks ago, and runs on your hardware rather than in a company's cloud. OpenClaw went viral, passed 100,000 GitHub stars in two months and started a debate about what AI assistants should be. The names Moldbot and Claudebot refer to the same project at earlier stages. Not a sponsored video, the project is free and open source.
ChatGPT, Claude, Gemini and Perplexity are powerful and useful, but they are not integrated into your workflow. You open a browser, type, wait, copy, paste elsewhere, repeat. Different tools for different jobs, none of which share context. Conversations are isolated, or context bleeds where you do not want it. And they do not live where you actually communicate, in WhatsApp, Slack or Telegram, but in a separate tab in a separate app.
An open-source, self-hosted assistant that runs entirely on your own infrastructure. Not cloud-based with local features, not download-our-app. It runs on your computer, a Raspberry Pi, a Mac Mini or a VPS. Instead of forcing another interface on you, it meets you where you already are: WhatsApp, Telegram, Discord, iMessage, Signal, Slack. You text it like a colleague and it answers in the same thread. Because it runs locally it has access to your actual tools.
Asked to list all markdown files in a demo directory, it returns the real files, not guesses. Asked to read a reading list and recommend which book to finish first based on progress, it opens the file, reads the content and reasons: finish the one already at 50 percent before the one at 25 percent. That is the difference between a chatbot and an agent with real access to a workspace.
OpenClaw is built on a gateway architecture: one process, multiple channels. It routes your message to whichever AI model you pick and integrates with over 100 services through the Model Context Protocol. Memory is persistent and context spans weeks. Because it is open source you can extend it, audit it and control every aspect of how it behaves.
Built by Peter Steinberger, an Austrian developer who sold his company and got bored. It started as Claudebot in November 2025, the name a combination of Claude and TARDIS, because like the time machine it is bigger on the inside. The mascot is a space lobster. Within 24 hours it had 9,000 GitHub stars, 60,000 by the end of the first month, over 100,000 today, making it one of the fastest growing repositories ever.
In my own setup it delivers summaries on topics I follow and prepares scripts, including groundwork for this video. The goal is automating my content pipeline so more time goes into the creative part and less into everything around it. Beyond that people use it for email, research, automation, smart home control and code generation.
Three reasons, even if you never run it. Privacy: your data stays local, no cloud logging, no tracking. Extensibility: you own the platform, the codebase is MIT licensed, you can change behaviour or add skills. Control: you decide which model, which APIs, which integrations and which machine it runs on.
IBM published an analysis arguing that OpenClaw challenges the assumption that AI agents must be vertically integrated. For years the industry insisted only large companies could build reliable agents because you need to control models, memory, tools, interface and security. Community-driven agents can work too, powerful, flexible and safe, if designed correctly. The trade-off is that you have to understand what you are doing.
OpenClaw is not for everyone. Security researchers from Cisco, 1Password and others raised serious concerns about prompt injection: manipulate the input and the assistant can be made to execute arbitrary commands. Researchers also found publicly accessible deployments with no authentication, exposing API keys, chat logs and full system access, because people followed instructions on X without understanding the setup they were creating.
Run it on a separate device, sandbox it, limit permissions, audit the code, or do not run it and watch the ecosystem instead. The project is still early: docs are improving, the community is growing, new skills appear daily in the extension marketplace. It is part of a wider shift where agents move from research papers into daily use.
The follow-up is the complete installation walkthrough: server, security, onboarding wizard and the first working message. Documentation is at openclaw.ai.
I work 1:1 with freelancers, consultants, coaches and small teams on practical AI workflows, including the question of what belongs in the cloud and what does not. First call is free, 15 minutes, no pitch.
Book a free intro call