What I learned building a local macOS MCP for ChatGPT-assisted coding/admin work
一位开发者构建了一个本地macOS MCP，旨在优化ChatGPT辅助的编码和管理工作流程，解决ChatGPT提供指令后仍需手动执行的痛点。该工具采用MIT许可，目前提供了81个MCP工具。开发者正在探索最佳的本地代理权限模型，包括按工具审批、配置文件或只读模式，以在确保安全性的同时，不影响工作流程的流畅性。
时间与来源
时间显示为 UTC
显示时区：UTC
本地时区尚不可用，暂时显示 UTC。
发布当时偏移：UTC+02026年9月9日 10:13 UTC
收录当时偏移：UTC+02026年9月9日 17:00 UTC
- 发布
- 2026年9月9日 10:13
- 收录
- 2026年9月9日 17:00
- 来源类型
- 开发者社区
- 档位
- 社区
- 信源状态
- 正常
档位是按信源手工设定的编辑判断，不是逐条打分。
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.