Building a DIY OpenClaw
What happened
It has been reported that a developer bought a Mac Mini and spent a weekend cobbling together a small, OpenClaw‑style setup so they could text with Anthropic’s Claude over iMessage. The project kept evolving, and allegedly the builder now uses the rig every day to chat, get help with tasks, and even have the agent write code. There’s a hint of glee in that detail — who wouldn’t want a polite, helpful assistant on their phone?
How it works
It has been reported that the Mac Mini runs Claude Code inside tmux; new iMessages are detected via imsg and piped into the tmux session. The agent is constrained by instructions: the only permitted way to reply is to invoke a ./send script, which calls imsg to text back. It also keeps a CONTEXT.md file for persistent notes and a logbook/ folder where short memories and links to full JSON transcripts are appended so the agent can dig through past conversations when needed. The builder also uses agent-browser to give Claude a browsing tool it can use for verification and web lookups while it codes.
Tools, tradeoffs and trust
It has been reported that credentials — including Claude’s API key — are stored in a 1Password vault that’s shared with the agent. That’s neat and practical, but it raises immediate privacy and security questions. The creator also wrote a suite of “skills” to encapsulate recurring tasks, effectively teaching the agent shortcuts and workflows. Clever, pragmatic, a little DIY mad scientist — and undeniably useful.
This is part of a broader trend: personal LLM agents that remember, act, and use tools. Is it thrilling or unnerving? Both. The heart of the story is simple: someone wanted a conversational assistant inside iMessage and built one. It feels a bit like having Jarvis in your pocket — minus the flying suit... for now.
Comments