← sources

I gave ChatGPT control of my Mac, then spent more time making its mistakes reversible than adding new tools

source: reddit (reddit:ChatGPT)published 2026-09-28 13:39ingested 2026-09-29 02:37

retained text

I maintain an open-source project called Mac MCP. I use normal ChatGPT as the orchestrator, while a local MCP server does the actual work on my Mac: browser tabs, files, shell commands, native UI, and delegated coding agents.

The first version was mostly about adding capabilities. The more useful work since then has been about making those capabilities boring enough to trust.

A few examples from the current 2.1.6 build:

- Safari and Chrome work can stay in background tabs with stable tab handles instead of constantly stealing focus.
- Foreground actions are capability-gated, so the model cannot simply decide to bring a tab/app in front of you.
- File mutations are journaled, and shell runs can produce whole-run filesystem undo receipts instead of leaving you to reconstruct what changed.
- Delegated coding agents can work in isolated Git worktrees and only apply changes back through an explicit conflict-aware step.
- If a mutating action has an uncertain outcome, Mac MCP refuses to automatically replay it. "Maybe it happened" is treated differently from "it failed."
- The local dashboard can show a sanitized "What Changed on My Mac" receipt without dumping prompts, commands, page content, or secrets.

That last part changed how I think about desktop agents. The interesting question is no longer "can ChatGPT click/type/run this?" It is "can I let it keep working without babysitting it, and still understand or reverse what it did?"

The project is free, MIT-licensed, and open source:
https://github.com/bulutarkan/mac-mcp

I am the maintainer. Curious where other people would draw the interruption boundary: what should always stop and ask you first, versus what is fine to do silently if it is scoped, visible, and reversible?

document fetched 2026-09-29 00:21 · canonical https://reddit.com/r/ChatGPT/comments/1wsewwu/i_gave_chatgpt_control_of_my_mac_then_spent_more