Codex App for Windows & COM Automation

Introduction

OpenAI have released the Codex Desktop App for Windows. It support WSL2 (Windows Subsystem for Linux), as well as running natively. As with the macOS app, Codex is positioned towards developers, but can actually do more, e.g. translating Excel files. As the Windows version uses PowerShell (and others), it can tap into COM-Automation, giving it first-party control over Microsoft Office content and actions - as opposed to open-source file format implementations that are sometimes lacking in accuracy. Other examples for COM components with potential uses for productivity is sending PDF files to a printer or processing pipeline through “AcroPDF.PDF” or using a database connection through “ADODB.Connection”.

In practice

The mode of operation is that the user first selects a working directory to operate in. They would then add files to operate on and issue a prompt - like in ChatGPT. A naive prompt to create a Hello-World Word document might look like this:

Create a Word file with the content: “Hello World”

However, Codex’ naive approach to this would cause Word to open a “File As” dialog for saving the file. If the user wouldn’t respond, Codex would conclude that process hangs - and proceed with a different method. What works instead is this - which could likely be added to the Skill library (textual procedural memory for the Agent):

Attention: when saving, you’ll have to handle the SaveAs dialog, so the process won’t hang.

This will create the document without any third party tools.

A more interesting example - translating a full Excel file - is in my post about Codex Desktop from a month ago:

Create an English version of Tabelle1_DE.xlsx in a separate file.

Note regarding Outlook: the “new Outlook” does not support COM (see “Outlook Object Model (OOM)” in this feature compatibility matrix).

Optimizing

Aside from adding Skills, the Codex release article notes that “Codex works best when a few common developer tools are already installed” - despite PowerShell being already powerful. Two noteworthy examples are Node.js or Python “to perform tasks more efficiently”.

Conclusion

As others have mused on social media if running Codex on macOS or Windows could be superior, I think the comparison is not so easy. Codex seems to make fewer mistakes on macOS (based on shell scripting + Python vs. PowerShell), but interfacing with productivity tools could open Codex to another user category - one different from engineers.