@dotey: I took some time to study the amontlabs/lcu project. It extracts Codex's Computer Use feature (the capability for AI to…

I took some time to study the amontlabs/lcu project. It extracts Codex’s Computer Use feature (the capability for AI to view screens, click mice, and type on keyboards to operate computers) as a standalone component, allowing other AI Agent tools like Claude Code, Codex CLI, Pi, and more to invoke it.

I originally thought it was an imitation of Codex that implemented its own system, but it turns out it’s directly using Codex’s Computer Use. It locates the locally installed ChatGPT desktop version, launches the built-in Computer Use runtime from OpenAI, and then connects it via MCP to the Agent tools you’re using. The repository documentation clearly states that instructions, execution, and permission policies all follow the OpenAI original runtime.

Therefore, the official ChatGPT desktop version must be installed on the computer, but there’s no need to log in to a ChatGPT or Codex account.

It provides very few tools to the model. Common computer control schemes would treat screenshots, clicks, inputs, and scrolling as separate tools each, but LCU only offers two: js and js_reset.

js is used to execute a segment of JavaScript code, with a preloaded object named cua in the runtime environment. The model operates the computer by writing code: first calling cua.getState() to retrieve the current desktop state and usage instructions, then using getApp to select a specific window, followed by invoking methods like click, typeText, pressKey, getScreenshot, and so on. Clicks, inputs, and result reads can all be written into the same code segment, completing everything in one call.

This JavaScript environment maintains state, so variables defined in one call can still be used in the next; if the environment encounters issues, js_reset can be called to clear and restart. Beyond screenshots, the model can also read accessibility information for windows (the list of UI elements provided by the operating system for screen readers), click on elements directly, and avoid complete reliance on screen coordinates.

The js plus cua object interface is OpenAI’s original design, forwarded as-is by LCU. Repository logs show that during comparative testing on macOS, the initialization instructions received by the model and the usage guide returned on the first call match the content received by the official Codex byte-for-byte.

Currently, the formally adapted Agent tools are Pi, Codex CLI, and Claude Code, with Oh My Pi and Hermes receiving experimental support. The system supports macOS on Apple Silicon chips (requiring accessibility and screen recording permissions), as well as Linux using X11 desktops (ARM64 and x86-64). Windows 11 is still in the candidate stage, and Intel Macs as well as Wayland desktops are not supported.

Controlling the Chrome browser requires additional setup, using OpenAI’s official Chrome extension, with each website still needing separate approval. Installation requires Python 3.12 or higher; the README suggests pasting a prompt to your own Agent to have it follow the installation docs for configuration, or you can manually download the installation package from GitHub.

Project address:

向阳乔木 (@vista8): 发现一个牛逼的东西，让任意 Agent 调用 Codex 的 Computer Use。

Codex 最强的就是 Computer Use。

但最近用 Claude Opus 5.5比较多，这样就强强联合了。

刚测试通过，安装后建议配置个 Hook，指定白名单可控制哪些 App

安装地址见评论区