Three weeks ago I bought a Mac Mini because of one YouTube video. Today that bot analyzes a hundred Telegram channels, sends me digests, and pushes code to my repo on its own.
The project is called OpenClaw (formerly Clawdbot, then Moltbot - Anthropic asked to rename it). Created by Peter Steinberger - founder of PSPDFKit, scaled it to a billion users, sold, retired. Got bored. Built an AI agent and released it for free.
Why I Bought a Mac Mini
Watched a few YouTube videos. Bloggers showed their bots building apps, creating PRs on GitHub, sending morning news briefings. The hype around the project was insane. I caved.
Mac Mini isn’t required - it runs on any hardware. But Apple Silicon with unified memory turned out to be ideal. Ordered.
Setup and First Days
One install command. Fast enough onboarding and initial config. Connected Claude through my existing Claude Max subscription as the brain. Created a Telegram bot. The whole setup took a couple of hours - I tried not to cut corners on security, since an agent with full computer access is serious.
Open source. Free. 50+ integrations and skills out of the box. Started exploring.
Then I gave it the first real task: build me a landing page. Fed it some description about me, recorded a couple of voice messages through Telegram, gave it links to LinkedIn, Telegram, YouTube. Done in 10 minutes, a couple of edits, deployed to Vercel. That’s the current prodfeat.ai landing.
Telegram Scraper + AI Digest
Then a bigger task. Built a Telegram Scraper - a thing that pulls posts from a hundred AI, startup, and product channels. Everything goes into a database.
Collecting is half the job. Reading it by hand - 2 hours of scrolling every day. Not sustainable.
Gave the bot database access. Now it analyzes hundreds of posts every night. Groups by topic, not by channel. Writes 2-3 sentence summaries. Sends me a ready digest in the morning.
15-20 minutes of reading instead of two hours. With links to originals. Twice a day.
Cost-wise: scraping is basically free, just server + database. All AI post-processing is covered by the Claude subscription. No extra token charges.
Shared Project - This Is What Surprised Me
We work through git. I push changes - it pulls. It adds its own - pushes back. Like a two-person team. Except one is a bot.
And it doesn’t just execute tasks. The bot suggested on its own how to improve the database structure. Updated SQL queries. Added its own working notes to analyze content better. Without being asked.
It marks post statuses: processed or not. Keeps its own notes to avoid duplicate analysis. Doesn’t re-read the same thing. Especially with images - every extra call costs both time and money.
How It Works Under the Hood
Three key mechanisms.
Memory. Stores everything in plain Markdown files. Daily logs memory/YYYY-MM-DD.md - like a journal. Remembers context between sessions. No need to explain who you are and what you’re doing every time. It knows.
Computer Use. Shell commands, browser, files. Essentially does what you’d do by hand. Through a sandbox.
Skills and MCP. Skills are ready-made behavior scenarios you can plug in and combine. MCP (Model Context Protocol) is a standard protocol for connecting to external APIs and tools. Together they give the agent extensibility: new capabilities and integrations without rewriting code. A similar principle is at work in custom Claude Code commands.
What I Learned in 3 Weeks
An AI agent isn’t “write me some text” or “answer a question.” It’s a full team member that can own its part of the work. You handle infrastructure - it handles analysis. Shared project.
But the most important thing is context management. Without it - chaos. With it - a predictable result every morning.
The project exploded. Over 170k stars on GitHub already. Cloudflare built their own adaptation for cloud deployment. The community is growing, there are ready-made Skills - from smart home to repo monitoring.
Sources
- OpenClaw GitHub - official repository
- Pragmatic Engineer: interview with the creator - Peter Steinberger on the project
- Memory Architecture Explained - technical breakdown of the memory system