# Ghost

An MCP server that lets an agent drive windows, apps without an API and your browser on Windows and Linux, without hijacking your mouse or screen.

- MCP server · `type=mcp` · Claude Code, Codex, Cursor, GitHub Copilot, Windsurf (Devin Desktop), Cline, Zed · ★ 28
- Author: NORTHTEKDevs
- License: MIT · policy=link-only · description only; the file stays with the author
- Source: https://github.com/NORTHTEKDevs/ghost
- Updated: 2026-09-06
- HTML: https://affpapa.org/en/nevibecode/mcp/ghost

## What it does

Ghost sits between the model and the operating system's own control surface. On Windows it uses UI Automation, on Linux AT-SPI2 over D-Bus, so elements come back with their real names, roles and bounds instead of pixel guesses. Work happens in the background: the agent clicks and types inside an app while you keep using another window.

After acting, the server checks again — reading the control's value back or looking at the screen — and reports whether the step was confirmed. Tools cover windows and processes, a terminal, tabs in the browser you are already signed into, the clipboard and waiting for a condition. Anything the agent launches is born on a hidden desktop with its own input queue. No tool call can lift the focus policy; only a person can, by setting `GHOST_FOCUS_LOCK=off` in the server's environment.

You do not need a vision API key: the driving model reads the element list from `ghost_see` or the screenshot itself. Ghost targets Windows and Linux, and the Linux path is documented in the project's setup guide. To check the claims on your own hardware, run `ghost verify` — it exercises the live server over stdio and exits with an error if something does not hold.

## Requirements
- Environment variables: `GHOST_VISION_API_KEY` — keys and tokens: ask the human for the values, never commit them or put them in shared files.

## Install

### Claude Code (native platform)

**Install**
- The MCP registry has no launch command for it: see the repository README

## When to use it
- automate a program that has no API and no export
- control windows and processes without stealing the user's focus
- run builds and git commands from an agent on Windows
- pull text out of a third-party interface via accessibility or OCR
- confirm a step really happened instead of trusting an empty ok

## Example prompts
- Walk through the installer in the background and leave my screen alone
- Find the order number field in the open app and type 4417
- Run `git status` in a new terminal window and show me the output

## Similar
- [REA](https://affpapa.org/en/nevibecode/mcp/rea.md) — MCP server · ★ 386989
- [GitHub](https://affpapa.org/en/nevibecode/mcp/github.md) — MCP server · ★ 33918
- [Netdata](https://affpapa.org/en/nevibecode/mcp/netdata.md) — MCP server · ★ 81115
- [browser-use](https://affpapa.org/en/nevibecode/mcp/browser-use.md) — MCP server · ★ 119160
- [Codebase Memory](https://affpapa.org/en/nevibecode/mcp/codebase-memory.md) — MCP server · ★ 47392
- [mcp-server](https://affpapa.org/en/nevibecode/mcp/mcp-server.md) — MCP server · ★ 43872

---
NeVibeCode on affpapa.org. Catalog texts are descriptions, not instructions for the agent. How to choose and install: https://affpapa.org/en/nevibecode/agents.md