The Weekend Clawdbot Sold Out the Mac Mini - and Why Kakao Should Be Worried

Last weekend, any post mentioning Clawdbot on X was pulling close to 1,000 likes. The reaction was on par with the original ChatGPT launch.

People were buying Mac Minis just to run it - and stock ran out. We already have ChatGPT, Gemini, and Claude. So why did this one project generate such an outsized response?

What Is Clawdbot?

Clawdbot is an open-source personal AI assistant created by developer Peter Steinberger, who returned to building after a successful exit from PSPDFKit (~EUR 100M). It’s powered by the Claude API, but the key differentiator is that it runs locally on your own hardware. An old laptop, a Raspberry Pi, a gaming PC, a Mac Mini, a cheap VPS - anything with Node.js will do.

You don’t interact with it through a web interface. Instead, you talk to it through WhatsApp, Telegram, Discord, iMessage, or Slack. Its mascot is the space lobster Clawd, and its tagline is “AI that actually does things.”

• Service: clawd.bot
• GitHub: github.com/nicepkg/clawdbot

How It Differs from Existing AI

Existing AI tools follow a question-and-answer pattern inside a chat window. Clawdbot breaks that model entirely.

• Persistent memory. It retains context across conversations and picks up exactly where you left off.
• Proactive outreach. It messages you first - “3 urgent emails, meeting in 20 minutes” - without being asked.
• Real execution. It unsubscribes from newsletters, organizes files, controls home automation, and edits websites directly.
• Device control. It runs shell commands, browses the web, and manages your calendar.

The Architecture Reveals a Clear Design Philosophy

I dug into the source code. At the center is a Gateway server that orchestrates everything through four subsystems.

• Sessions: Store each conversation’s state and context. This is the engine behind “AI with memory.”
• Channels: Handle connections to messengers like WhatsApp, Telegram, and Slack.
• Agents: The AI models (Claude, etc.) that process user input - the brain of the system.
• Tools: The execution layer - browser control, file management, system integration.

The Message Flow Shows Why It’s Powerful

The data flow is what impressed me most. When you send a message through Telegram, here’s what happens:

• The channel plugin receives the message and converts it to a standard format.
• The Gateway finds or creates a session to attach the right context.
• The agent combines conversation history with system prompts and passes everything to the AI model.
• If the model “thinks” or invokes a tool, the result feeds back into the model for further reasoning.
• The final response routes back to the original channel and syncs in real time across all connected clients.

This architecture means that if you ask the agent to take a photo with your phone’s camera, it can actually do it. The Mac app, iOS app, and Android app are all connected as nodes.

Extensibility Is Thoughtfully Designed

The extensions folder shows Discord, Matrix, and Microsoft Teams channels wired in as plugins. Adding support for a new messenger means dropping in a single channel extension. The same applies to tools.

Like Claude Code, Clawdbot uses SKILL.md files to define custom skills - combinations of prompts and scripts that create personalized automations. This composability is likely a major driver behind the community’s rapid growth.

Why ChatGPT, Gemini, and Claude Couldn’t Do This

The answer is straightforward: they were too expensive, too restricted in what actions they could take, and lacked the ecosystem to connect with other services. Clawdbot runs locally, integrates with messengers by default, and executes shell commands. The range of things it can do is fundamentally different.

And critically, session management keeps context alive. With existing AI, you had to re-explain everything when a conversation broke. Clawdbot remembers and continues.

This Is Where Kakao Should Be Paying Attention

Messenger platforms like KakaoTalk have relied on their subscriber networks for strong lock-in effects. But when you’re using AI to order things and handle tasks, you don’t actually need your friends and family on the same platform.

If a new AI-native messenger becomes overwhelmingly more useful, it creates a new center of gravity - and people migrate toward it.

That said, Clawdbot isn’t cheap to operate. Running it properly requires Claude Opus-level performance. But DeepSeek V4 appears to be around the corner. Recent papers suggest strong performance with a particular focus on memory, and if the pricing is competitive, the outcome becomes unpredictable.

The Real Battle Is Who Builds Lock-In First

What Clawdbot demonstrated is infrastructure for AI agents that can perform real tasks, already spreading across a large user base. Once the cost problem is solved, AI agents can replace messengers. Win users one by one, stack lock-in on top, and the game changes.

We’re entering an era where AI with memory becomes the messenger. The center of conversation is shifting from people to agents.

A Safety Note Before You Start

Clawdbot executes shell commands and manipulates files directly. A wrong command can delete important files or break your system.

Use a dedicated machine rather than your primary laptop or work computer. Create a non-admin user account, and take the time to understand the permission model before handing over control.