Back to browse
GitHub Repository

Full computer-use for AI agents. Self-learning workflows. Native macOS. No screenshots required.

1,504 starsSwift

Ghost OS – Let AI agents use your Mac, not just the terminal

by mcheemaa·Feb 19, 2026·1 point·0 comments

AI Analysis

●●●BangerZero to OneBig BrainSolve My Problem

Accessibility tree instead of screenshots: structured, fast, works with any native macOS app.

Strengths
  • Clever architecture: reads macOS accessibility tree instead of vision-guessing pixels—orders of magnitude faster and more reliable.
  • Recipes as learnable JSON workflows: frontier model figures it out once, small models replay forever—real cost/capability win.
  • Native app support: not just browsers. Slack, Finder, Mail—anything with accessibility labels. Genuine differentiation.
Weaknesses
  • macOS-only today. Windows/Linux users and non-accessibility-compliant apps are out of luck.
  • Early project (v2.0.4, 6 stars). Recipe ecosystem thin; most value locked behind frontier model usage to generate new workflows.
Target Audience

AI agent developers, Claude Code / Cursor users on macOS, AI automation enthusiasts

Similar To

Anthropic Computer Use · OpenAI Operator · Cursor's agent mode

Post Description

I use Claude Code every day. It can write code, run tests, search files, but it can't click a button, read what's on screen, or send an email. It's stuck in the terminal.

Ghost OS fixes this. It's an MCP server that gives any AI agent access to every app on your Mac, not just browsers. Native apps, Electron apps, menu bars, Finder, everything.

Most computer-use tools rely on screenshots and pixel guessing, or require Docker containers and sandboxed browsers. Ghost OS takes a different approach: it reads the macOS accessibility tree, which the OS provides for every app. Structured data about every button, text field, and link. Faster, lighter, and works with apps that never touch a browser.

When the AI figures out a workflow (send a Gmail, post on Slack, download a paper), it saves it as a JSON recipe. Plain text, auditable, version controllable. A frontier model learns it once, a small model replays it forever.

brew install ghostwright/ghost-os/ghost-os && ghost setup

Two commands. No Docker, no sandboxes, no container overhead. Works with Claude Code, but the MCP protocol means any agent can plug in. MIT licensed, written in Swift. Built on steipete's AXorcist for the accessibility layer.

Happy to answer questions.

Similar Projects

Automate Mac with Codex: macOS Control MCP Demo

Lets agents actually see the screen and act on it by returning OCR text with pixel coordinates and offering commands like click_at, type_text, and press_key. You can run it instantly with npx (it auto-creates a Python venv and hooks into Apple Vision/Quartz), and there are ready-made integration snippets for Claude, VS Code, and Cursor — a pragmatic, technically neat tool for closed-loop agent UI work. It’s limited to macOS 13+ and Apple APIs, but within that niche it removes a lot of friction.

WizardryNiche Gem
peterhddcode
104mo ago
Developer Tools●●Solid

Metatron – give coding agents your team's conventions over MCP

Structured decision records beat static .cursorrules files for maintaining team consistency.

Solve My ProblemBig BrainShip It
kerbelp
106d ago