Simon Willison reported on September 1, 2026 that OpenAI's Codex desktop app -- since rebranded to ChatGPT -- keeps 1.7 GB of bundled software in the user's cache folder, in a directory called codex-primary-runtime. Inside are a full Python installation, a full Node.js installation, and native binaries for Poppler, git, and the complete LibreOffice office suite. A sibling folder holds skills that tell Codex how to find and use each of them. The post reached 257 points and 120 comments on Hacker News, one of the day's most-discussed items.
Key facts
- 1.7 GB total in
~/.cache/codex-runtimes/codex-primary-runtime, including full Python and Node.js installations plus git, Poppler, and LibreOffice binaries. - The bundled skills live under a
plugins/openai-primary-runtime/plugins/documentsfolder and instruct the agent on how to invoke each binary. - Found September 1, 2026 by Simon Willison, using the disk-usage tool OmniDiskSweeper -- not documented or announced by OpenAI.
There is a straightforward reason an AI app would ship an office suite, and it is worth stating before the criticism. If a user asks an agent to read a spreadsheet, edit a presentation, or convert a document to PDF, the agent needs software that actually understands those formats. Parsing a modern office document correctly is a decade-scale engineering problem that LibreOffice and its OpenOffice ancestor have been working on since 2010. Reimplementing it would be foolish. Shelling out to it is the boring, correct answer.
The same logic covers the rest of the manifest. Poppler renders and extracts text from PDFs. Git handles version control. Python and Node.js are the two runtimes that most generated code expects to exist. What the bundle really describes is the agent's hands -- the set of real programs it can reach for when text generation alone will not do the job. Our explainer on tool use and function calling covers the general pattern; this is that pattern at its most literal.
The interesting shift is architectural. Two years ago, an AI product was a text box in front of a model. What Willison found is a full local execution environment with an office suite in it, shipped quietly as an implementation detail. The model is one component of a desktop application that also happens to include most of a Linux userland. That inversion -- where the model shrinks to a subsystem inside a large conventional program -- is the actual news here, and it is a much better predictor of where agent products are heading than any benchmark.
The reasonable objection is about consent rather than size. This is not documented. Nobody installing a coding assistant expects it to place LibreOffice on their machine, and Willison found it by running a disk-usage tool out of curiosity, not by reading release notes. There is a real difference between "the app needs these dependencies" and "the app silently installs 1.7 GB of third-party software into a cache directory where users will never look." The distinction matters for anyone maintaining a fleet of machines: bundled binaries have their own vulnerability histories and their own patch cadence, and a security team that does not know LibreOffice is on the endpoint cannot patch it there.
The bloat complaint is the weakest one, though it is the one that generated the most comments. Cached runtimes are recoverable disk, not resident memory, and 1.7 GB is not much on a modern machine. The bigger question is whether it stays 1.7 GB. Every capability an agent gains tends to arrive as another bundled binary, and cache directories are famously where software goes to accumulate. This is the same shape as the observation that the ChatGPT/Codex app's ambitions now extend well past code -- an agent that does office work needs office software, and there is no principled place for that expansion to stop.
The honest caveat: this is one developer's inspection of one machine's cache, published as a short note rather than an investigation. OpenAI has not commented, and the exact contents of that folder likely vary by platform, app version, and which skills a user has actually exercised. Willison's screenshot and directory paths are specific and checkable by anyone with the app installed, which is the right standard for a finding this small. But nothing here establishes that OpenAI is doing anything improper -- only that it is doing something undocumented, and that the shape of an AI desktop app in late 2026 looks a lot more like a conventional software distribution than the marketing suggests.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)