convez.ai
MCP ServerAutomaçãoIntermediário
opieeipo

MCP Window Vision

opieeipo/MCPWindowVision

Visão de tela + logs Unity para assistentes IA em macOS/Windows/Linux

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP multiplataforma que permite a assistentes IA visualizar a tela e acessar logs do Unity Console. Perfeito para agentes trabalhando em aplicações GUI (como projetos Unity) que precisam ver resultados e ler erros diretamente, sem depender de screenshots manuais.

Instalação

$npx convez-ai opieeipo/MCPWindowVision

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
v0.2.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Agentes IA debugando projetos Unity automaticamente
  • Assistentes com visão de tela para tarefas de automação GUI
  • Leitura em tempo real de erros e logs do Unity Editor

Veredito

Ferramenta sólida e sem dependências externas complexas para dar visão computacional a agentes IA. Especialmente valiosa para fluxos de desenvolvimento Unity/games onde feedback visual é crítico; implementação multiplataforma robusta com suporte integrado a logs do editor.

README

A tradução para o português deste README ainda está na fila.

MCP Window Vision

A small, cross-platform MCP server (macOS / Windows / Linux) that gives an AI assistant eyes on your screen and access to Unity's Console log — so an agent working on a GUI app (like a Unity project) can see the result and read the errors itself, instead of asking you to screenshot and paste.

Screenshots use mss + Pillow (both cross-platform). The MCP transport is pure standard library. Runs on Python 3.9+.

Tools

ToolWhat it does
capture_screenScreenshot the main display (downscaled), returned as a PNG image
capture_regionScreenshot a rectangle {x, y, w, h} (pixels)
read_unity_consoleTail Unity's Editor.log (Debug.Log / errors / compile messages); optional substring filter

Requirements

  • macOS, Windows, or Linux (Linux capture needs X11)
  • Python 3.9+
  • macOS only: Screen Recording permission for whatever process launches the server (your terminal / MCP client): System Settings → Privacy & Security → Screen Recording.

Unity Editor.log is found automatically per-OS:

OSPath
macOS~/Library/Logs/Unity/Editor.log
Windows%LOCALAPPDATA%\Unity\Editor\Editor.log
Linux~/.config/unity3d/Editor.log

Install

Recommended — pipx (cross-platform; isolates the tool in its own environment and puts it on your PATH, sidestepping PEP 668 "externally-managed-environment" errors from system/Homebrew Python):

pipx install mcp-window-vision

Don't have pipx yet? Install it first, then re-run the above:

brew install pipx      # macOS / Linux
scoop install pipx     # Windows
pipx ensurepath
python3 -m venv ~/.venvs/mcp-window-vision
~/.venvs/mcp-window-vision/bin/pip install mcp-window-vision

Then use the absolute path ~/.venvs/mcp-window-vision/bin/mcp-window-vision when registering below.

Verify capture works (and the OS permission is granted):

mcp-window-vision --selftest
# -> "capture OK — NNNNN base64 chars on <OS>"

Use with Claude Code

User scope (available in every project):

claude mcp add window-vision --scope user -- mcp-window-vision

…or per-project via a .mcp.json at the project root:

{
  "mcpServers": {
    "window-vision": {
      "type": "stdio",
      "command": "mcp-window-vision"
    }
  }
}

If mcp-window-vision isn't on your PATH (e.g. the plain-venv install above), use its absolute path as the command. Restart the session so the client picks up the server.

Develop

git clone https://github.com/opieeipo/MCPWindowVision
cd MCPWindowVision
pip install -e .
mcp-window-vision --selftest

Roadmap

  • v0.1: macOS screen/region capture + Unity log reading.
  • v0.2 (current): cross-platform (macOS / Windows / Linux) via mss + Pillow.
  • next: a Unity-side C# bridge for live editor control (run menu items, inspect serialized references, enter/exit Play) — closing the loop fully.

License

MIT

Ver o original no GitHub

Bom para

Relacionados