Codex Computer Use for Windows

Give Codex computer-use on Windows with any third-party model — no official OpenAI computer use required.

Codex Computer Use for Windows is a computer-use layer for Codex users who run third-party models (DeepSeek, GPT-5.6, Kimi, or any OpenAI-compatible API). If you use Codex with a custom model provider, the official Codex computer-use feature is unavailable — this project gives you that capability instead.

Chinese version: README.zh-CN.md

How it works

• Codex (running any model with tool/function calling) decides the next action.
• MCP server (windows-mcp, MIT) exposes Windows automation tools (click, type, scroll, screenshot, UI state, ...).
• Windows UI Automation (UIA) reads the screen as text + coordinates, so no vision model is needed — text-only models like deepseek-v4-flash work out of the box.

You → Codex (DeepSeek / GPT-5.6 / Kimi / ...) → MCP tools → Windows UIA → click/type/scroll

Quick start (new Windows VM)

Option A — one-line bootstrap:

powershell -ExecutionPolicy Bypass -c "irm https://raw.githubusercontent.com/ukinch605/codex-computer-use-windows/main/scripts/install.ps1 | iex"

Option B — manual:

git clone https://github.com/ukinch605/codex-computer-use-windows.git
cd codex-computer-use-windows
powershell -ExecutionPolicy Bypass -File setup.ps1

Then:

1. Install the MCP server into Codex:

   powershell -ExecutionPolicy Bypass -File install-codex-mcp.ps1

2. Restart Codex (the MCP server is loaded when a session starts).

3. Just ask Codex to operate the computer, e.g.:

   打开记事本，输入 hello world

   用 Edge 打开 example.com，把页面标题告诉我

4. Optional verification:

   python scripts\mcp_smoke_test.py        # MCP server starts and config fragment is valid

Supported models

Model	Status	Notes
DeepSeek (deepseek-v4-flash / deepseek-v4-pro)	✅ Tested	https://api.deepseek.com, tool calling works
GPT-5.6 / Kimi / other OpenAI-compatible models	⚠️ Expected	Should work via any OpenAI-compatible endpoint; not individually tested yet

Any model that supports OpenAI-style tool/function calling should work. Follow your provider's docs for base_url and model name.

Optional: standalone agent mode (without Codex)

If you don't use Codex, the same Windows automation stack can run as a standalone agent driven directly by a model API (Windows-Use based):

python scripts\agent\smoke_test.py   # API connectivity + function calling
python scripts\agent\gui_test.py     # Notepad: type "hello world"
python scripts\agent\edge_test.py    # Edge: open example.com and read the title

Or double-click 启动Windows-Use.cmd.

Test matrix

See docs/TEST_MATRIX.md for results recorded on real virtual machines.

Troubleshooting

See docs/troubleshooting.md.

Contributing

See CONTRIBUTING.md.

License

MIT. This project integrates Windows-MCP (MIT) and Windows-Use (MIT).