What I learned building a local macOS MCP for ChatGPT-assisted coding/admin work
A developer created a local macOS MCP to streamline ChatGPT-assisted coding and admin tasks, addressing the inefficiency of manually executing commands or edits suggested by ChatGPT. The MIT-licensed tool currently offers 81 MCP tools. The developer is seeking input on effective local-agent permission models, considering options like per-tool approvals, profiles, or read-only modes, to balance security with a smooth workflow.
Time & source
- Published
- 09/09, 10:13 UTC+0
- Ingested
- 09/09, 17:00 UTC+0
- Source type
- Dev community
- Tier
- Community
- Source status
- Healthy
Tier is a per-source editorial setting, not a per-item score.
I kept running into the same workflow problem: ChatGPT could tell me exactly what command to run, what file to edit, or what to click in Safari, but I still had to leave the chat and do the mechanical part myself.
Closest alternatives I tried were separate shell MCPs, browser tools, or just using Codex/Work for heavier jobs. Those are useful, but I wanted the normal Chat surface to be able to carry a local task further without turning my Mac into one giant unrestricted shell.
So I built Mac MCP. It is a local macOS MCP server with shell/files, Safari/Chrome automation, macOS Accessibility actions, persistent memory, Agent Skills, delegated Codex/OpenCode workers, voice input, and a small native menu-bar controller/dashboard.
The useful bits I learned while building it:
- Stable browser tab handles matter a lot. Background Safari work becomes much less annoying when an agent can inspect one tab without constantly stealing focus.
- Local permissions need to be obvious. The dashboard is localhost-only, auth can be required, and dangerous capabilities are explicit rather than hidden behind a generic “computer use” button.
- Delegating long sub-tasks to Codex/OpenCode works well, while the main Chat thread stays the coordinator.
- Chat, Codex and Work are different surfaces; this project is mainly about making the regular Chat side more agentic for longer multi-step local tasks rather than pretending they are the same thing.
It is MIT licensed and currently exposes 81 MCP tools. I’m still figuring out the best permission model (per-tool approvals vs profiles vs read-only modes).
Repo: https://github.com/bulutarkan/mac-mcp
Would be interested in how other people structure local-agent permissions without making the workflow painful.