Openclaw

Openclaw is my personal AI agent framework. I built it on my own PC so it could run local models like Qwen Coder 2.5 alongside cloud calls to Claude and GLM when a job needed the upgrade. The whole thing lives on a Pop!_OS Linux install I dual-booted off my Windows machine, so the agent has its own isolated environment to operate in without sharing a session with the rest of my computer.

Eventually I migrated openclaw to its forever home: an older i7 Dell XPS 13 that sits on my desk doing one job. That's where it runs now, on its own hardware, on its own OS, with its own set of credentials.

Today openclaw posts to LinkedIn for me, fully manages my MushroomPro WordPress site (that project is openclaw's main public case study), and runs as a personal note-taking and feedback-loop system on the parts of my business I need to stay close to. It pulls in updates, surfaces what's drifted, and helps me catch the things I'd otherwise miss between weekly check-ins.

Most of what I know about agent design (where they break, where they're worth the credit spend, where they need a human in the loop) came from running this thing every day. I wrote up what I learned in the Minimax vs GLM retrospective and the Gemini vs Claude review.