Everyone’s installing OpenClaw right now, and after using it for a few days I realized: this is basically an agent running on your system and doing stuff for you. The biggest concern I kept hearing was privacy—so I decided to reverse engineer the idea and build my own OpenClaw-style agent that runs locally for my use case. In this video I demo both a CLI (terminal/TUI) and a simple UI, with bidirectional sync so whatever you type in the terminal shows up in the UI too. I also added settings like memory options, LLM config (I’m using GPT-4), and cron jobs. Then I walk through what the agent can actually do: check CPU usage, run shell commands (like checking Node version), do web search, and perform file operations (read/write) inside a dedicated sandbox/workspace folder so it doesn’t touch random parts of my system. I also show a full “build” flow—generate a simple HTML site for a pet shop, store it in a workspace, host it locally, then run Git operations like init and commit. Finally, I break down how I built it: persistent memory stored as JSON session logs (no DB), Python-based “skills” (file, terminal, git, web), Playwright + a search API for web browsing/search, and a tool registry that defines what the agent can do. I also used LiteLLM (the AI picked it over LangChain) and wrapped everything with a FastAPI backend + a basic index.html frontend. Next step: Telegram access and running it on VPS/Azure.