📌 TL;DR

• Clawdbot (now Moltbot) → Everyone's losing their minds over this always-on AI assistant. It's cool in theory, but has limitations. If you're not already using Claude Code daily, this solves a problem you don't have yet.

• Claude tool integrations → Figma, Slack, Asana, Canva now render as live UIs inside your Claude chat.

• Skills.sh → A library of pre-built skills you can install into Claude Code or Cursor. Worth browsing for marketing, design, and dev shortcuts.

• Kimi K2.5 → New open-source model from Moonshot beating Opus 4.5 and GPT-5.2 on several benchmarks. Free at kimi.com.

• Vibe branding → Vibe branding my new project using Midjourney + a custom Claude skill for prompting. Free skill link inside.

Clawdbot (now renamed to moltbot)

If you've been anywhere near AI Twitter this week, you've seen Clawdbot. "The most powerful AI tool I've ever seen." "What Siri should have been." "The Jarvis era is here."

I haven't tried it yet myself, but I've spent the week diving deep into what it actually is, what it can do, and whether you should care. Here's my honest take…

What Clawdbot actually is

Clawdbot (now renamed to Moltbot after Anthropic's lawyers got involved) is an opensource tool built by Peter Steinberger - not affiliated with Anthropic or Claude in any way.

It lets you run an AI assistant that's always on. Instead of opening a terminal or chat window, you message it through Telegram, WhatsApp, Slack, Discord, or iMessage - from your phone, while you're out, wherever.

It runs 24/7 on your own hardware (a Mac Mini, a VPS, even a Raspberry Pi), maintains persistent memory across days and weeks, and here's the genuinely new part; it can message you first. Set it to send you a morning briefing, alert you when something urgent lands in your inbox, or prep your calendar overnight before you wake up.

That proactive, always-on nature is legitimately new & cool. I saw one guy on X connect it to his 8 sleep, so it knows when he wakes up, and orders him breakfast on doordash.

It's the first real step toward the "AI assistant that actually assists" vision we've been promised for years.

But ...

Most of the viral demos are underwhelming.

I've been watching these all week: twitter research, organising a downloads folder, sending good morning texts to your wife, summarizing group chats. One guy proudly showed Clawdbot sorting files by type - something Finder does with a single click.

These aren't insane paradigm shifts. This is stuff Claude Code already does. The difference is you're triggering it from Telegram instead of a terminal.

Nick Saraev put it bluntly: "Claudebot is Opus 4.5 wrapped around Telegram, plus cron. That's it." (cron is just a time-based job scheduler)

He's not entirely wrong. The capabilities are identical to Claude Code. The difference is when and where you can trigger them.

If you're not already using Claude Code to run AI employees for marketing, content, research, automation, then Clawdbot is solving a problem you don't have yet.

The founder himself says most people shouldn't install it

Direct quote from Peter Steinberger, Clawdbot's creator: "Most non-techies should not install this. It's not finished. I know about the sharp edges."

He's not kidding.

You most definitely can set Clawdbot up securely. But most people aren't doing that. They've seen a few viral tweets, rushed to install it, and left their servers wide open.

To run Clawdbot on a server, you need to put it on a publicly available URL. Problem is, there are services that constantly scan and index every URL on the internet. Security researchers searched "Clawdbot Control" and found thousands of servers they could access with a single click - many without password protection. Full access to people's messages, connected services, and API keys.

One researcher found a random user, fully identified them, and pulled their entire message history. Another extracted someone's crypto private key in five minutes using a single email that tricked the AI into handing it over.

And even when it's set up correctly, it's expensive. Unlike a $20/month subscription, you pay for every task the AI runs. Matthew Berman spent $130 in a single day. Others hit $300+ in 48 hours doing what they thought were basic tasks.

The bottom line

Don’t get me wrong Clawdbot points toward something genuinely exciting. An AI that works proactively, maintains long-term memory, and lives where you already communicate. That vision is real.

But don't get FOMO’d into something you don't need yet.

If you're not already using Claude Code, or more advanced agent builders daily for content, marketing, research, automation, etc… Clawdbot isn't going to change your life. You'd be adding complexity on top of a foundation that doesn't exist yet.

Get good at Claude Code first. That's the unlock. Clawdbot can wait.

Also this week…

Claude now opens your work tools directly inside the conversation. Figma, Slack, Asana, Canva, and more now render as live, interactive UIs right in your claude chat.

Vercel launched skills.sh, a library of pre-built skills you can add to AI tools like Claude Code or Cursor. Think of it like downloading expertise. Install a skill, and your AI just knows how to do that thing. It’s worth browsing. There's skills for marketing (copywriting, SEO audits, launch strategy), web design, video creation, and more.

Kimi K2.5 dropped. A new open-source model from Chinese AI company Moonshot that's beating Claude Opus 4.5 and GPT-5.2 on several benchmarks. It's free to use at kimi.com, and the weights are on Hugging Face if you want to run it locally. Also includes "Kimi Code" - their answer to Claude Code.

💡 Builder’s notes

Nothing too crazy to show you this week. I've been deep in building Vibe Founder, my new AI education company.

I’m currently vibe branding this project, and i'm going with this modern renaissance theme. The core mission idea behind Vibe Founder is articulated beautifully in my manifesto, and I think that the renaissance encapsulates the core idea nicely - the artisan returns.

I used Midjourney (an AI image model) for these shots.

A general rule of thumb is that Midjourney creates art; GPT and NanoBanana deliver assets. To help with prompting Midjourney I created a Claude skill, to teach claude about how to prompt Midjourney. you can steal it for free here.

I combined 4 different —sref codes in Midjourney to nail this futuristic renaissance style that I can use in any scene.

stay tuned. waitlist launching soon 👀

🥣 Brain food

• Claude Code might kill n8n → Great example of how Claude Code can automate a task like competitor research that in the past would have been an n8n workflow. I'm genuinely starting to think traditional automation tools are on borrowed time.

• Ben Affleck & Matt Damon on AI's limits → I touched on this exact premise in my manifesto. AI cannot have the unique lived experience and creative insights that make us human. It won't replace creative work, only aid it.

• Living in -70°C → Mainstone visits Siberian communities who live in temperatures colder than the average on Mars. I love this guy's stuff. Makes you grateful for central heating and puts your "problems" in perspective.