I built an MCP server that lets Claude Code control your desktop

Two months ago I typed "open chrome and find latest news" into Claude Code and got:

I can't open Chrome or control your browser. I'm a code assistant, I work with files, run shell commands, and help with software engineering tasks.

Today, that same Claude books flights for me, fills out forms, checks competitor pricing every morning at 9am, and triages my email before I sit down at my desk.

The difference is an MCP server called OpenOwl (adds screenshot, click, type, OCR tools to Claude) plus a specific prompting pattern I figured out through trial and error.

The core problem: if you ask Claude to do a complex desktop task right away, it refuses. Too many unknowns. But if you build up gradually, it works every time:

1. "Take a screenshot of my desktop" < Claude discovers it has eyes

2. "Open Chrome" < Claude discovers it can interact

3. "Open Chrome and search for flights to London" < small mission

4. Repeat 2-3 times, then give it the full task

5. Schedule it to run daily

The progression matters. Skip to step 4 and you get refusals. Build up to it and Claude figures out its own capabilities.

I wrote up the full method with exact prompts, what I automate daily, honest limitations, and setup instructions: https://openowl.dev/blog/how-to-make-claude-control-your-computer