Simon Willison found the Codex desktop app bundles 1.7 GB, including LibreOffice: Reveneau

Why it mattersA coding agent that includes its own office suite and document toolchain can convert and read files locally, so file handling stops being an API call and becomes disk space on the developer's machine.

Simon Willison was trying to free disk space on his Mac on 1 September and found 1.7 GB in the cache folder used by OpenAI's Codex desktop app, the one that has since been rebranded to ChatGPT. The folder does not hold model weights. It holds complete working copies of other people's software.

What is in the folder

Willison published the breakdown from his disk usage tool. Native binaries account for 771.0 MB of the total. Node.js takes 446.4 MB and Python takes 440.6 MB. Inside the native binaries is a headless build of LibreOffice at 429.7 MB, Poppler at 187.9 MB, git at 148.1 MB, libheif at 4.7 MB and jxrlib at 679.9 kB.

LibreOffice is the one that stands out. It is a full office suite, and the headless build exists to convert and read documents without ever opening a window.

Why it is there

Willison points to a documents plugin folder inside the same runtime directory. In his words, it "includes skills which tell Codex how to find and use those binaries." So when the agent opens a spreadsheet or a PDF, it uses a local toolchain and a set of instructions for running it, with no hosted conversion service.

Willison reports the finding and the folder contents. He does not say why OpenAI made this choice, and OpenAI has not published an explanation of the bundle, so the reasoning behind it is not established.

Bundling the toolchain instead of calling a service

This is worth knowing for two reasons, and neither is the disk space.

The first is that the design of agent products is changing. Reading a PDF or a Word document used to mean an upload and a hosted parser. Including Poppler and LibreOffice on the client puts that work on the user's machine, which keeps the file local and removes one network request from a step agents perform constantly. If you are designing a tool that handles customer documents, that is different for privacy from a hosted pipeline, and it is a real design option rather than a theoretical one.

The second is the cost side of the same decision. A 1.7 GB runtime cache is a large download and uses the same space on disk once installed, and it grows with every capability added this way. A team deciding whether to bundle a large dependency or call a service now has a visible example of a major vendor choosing the bundle, along with the number that choice costs on disk.

There is also a plain practical note for anyone short on space. The folder is a cache, and it sits under the user's home directory rather than in the application bundle, so it is easy to miss when auditing what an installed tool actually occupies.

Source

Simon Willison, "The ChatGPT/Codex app bundles a full copy of LibreOffice"

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.