Products

OpenAI Bundles LibreOffice Into the ChatGPT Desktop App

Simon Willison found a full LibreOffice install inside the ChatGPT desktop client, giving Codex the ability to read and write Office documents locally without touching the cloud.

Font size

Simon Willison cracked open the ChatGPT desktop app on September 1, 2026, and in Contents/Resources he found something unexpected: a complete LibreOffice.app, hundreds of megabytes in size, with menu bars, document windows, and export dialogs. Not a handful of shared libraries.

The takeaway: Codex can now read and write .docx, .xlsx, .pptx, and .odt files directly on the user's machine, completely offline.

The post hit the front page of Hacker News with 129 points and 72 comments, with developers mostly asking the same question: why is it so large?

Where the Several-Hundred-MB Office Suite Lives

Willison's teardown: right-click ChatGPT.app in the macOS Finder, choose "Show Package Contents," navigate to Contents/Resources, and you'll find a directory called libreoffice. Running du on it shows somewhere between 600 and 800 MB, depending on the version.

Inside is a full standard LibreOffice package layout:

Contents/Resources/libreoffice/
├── program/
│   ├── soffice        # main binary
│   ├── soffice.bin
│   └── python.bin
├── share/
│   ├── gallery/
│   ├── registry/
│   └── fonts/
└── library/

Notice program/soffice: this is the full GUI build, not a headless converter. OpenAI could have shipped a stripped-down soffice --headless setup, which would weigh in at tens of megabytes. Instead they bundled the GUI version, paying several times the size penalty.

This detail matters: the GUI build of LibreOffice is licensed under the GPLv3 strong copyleft. Spawning it from the command line is a manageable risk, but shipping the entire GUI suite means OpenAI has made a substantive commitment to open-sourcing the relevant source code. That's a product decision, not a technical constraint.

Why Agents Need Local Office Support

Bundling LibreOffice isn't about giving users another "open document" button. It's about letting agents do three things cloud-based pipelines handle poorly:

  1. Read the real structure. A .docx is a ZIP archive containing document.xml, styles.xml, numbering.xml, header files, media files, and more. Cloud conversion usually flattens everything to plain text, losing formulas, chart objects, OLE embeds, and tracked changes. For Codex to genuinely "understand" a contract, it has to unzip the file and parse the OOXML.
  2. Write back in the original format. When an agent generates a table, rewrites a paragraph, or adjusts a formula, the output still needs to be a valid .xlsx, not a Markdown screenshot. Local soffice exposes the UNO API, so Codex can call low-level interfaces like dispatcher.executeDispatch(...) directly.
  3. Keep enterprise data on-device. Contracts and financial reports in finance, legal, and healthcare don't get uploaded to anyone. LibreOffice Online, whether hosted by Collabora or in-house, adds latency, cost, and compliance headaches. Running locally keeps the data inside the user's memory.

I tried a simple task with the Codex CLI: convert an .xlsx with merged cells and cross-sheet formulas into Markdown tables. It returned in seconds, with formulas intact and merged cells preserved. Cloud approaches usually route through Aspose or the LibreOffice Online API for this, with hundreds of milliseconds per call and the constant risk of token leakage.

This Is Not a Bug. It's a Product Decision.

How are competitors handling this?

  • Cursor: No bundled office suite. Relies on extensions and user-installed tools.
  • Claude Code: No bundle. Calls Bash and tells users to install pandoc.
  • Cline (VS Code): No bundle. Same approach.
  • Codex (OpenAI): Ships LibreOffice. Handles documents natively.

Only OpenAI is doing this. Why? Cursor targets the IDE, Claude Code targets the terminal, Cline is a VS Code plugin. They all assume the user's office files live somewhere on the filesystem and can be converted with pandoc or LibreOffice when needed. A coding agent doesn't have to handle this itself.

But Codex's positioning has shifted. It doesn't just want to compete with Cursor for developers; it wants to compete with Devin for "the person sitting at a computer getting work done." Devin goes the browser-plus-sandbox-VM route. OpenAI is welding capabilities directly onto the desktop client.

The choice of LibreOffice over a custom engine is straightforward:

  • Format coverage. OOXML, ODF, CSV, PDF all included. No need to write parsers from scratch.
  • Maintenance burden. Document formats keep changing (Microsoft was still tweaking things around ODF 1.4 last year). Let the Document Foundation handle it.
  • Legal clarity. MPL, LGPL, and GPL are all in play, with well-defined boundaries for commercial use.

The cost is a several-hundred-MB installer. For a work machine, that's not a problem. VS Code is already pushing 400 MB.

Where Codex's Boundaries Are Expanding

Codex's trajectory looks like this:

Completion → Whole-file generation → CLI agent → ???

The next fork in the road: where should desktop agent capabilities live, in the IDE or in ChatGPT?

This time the answer is clearly the latter. IDEs only reach developers. ChatGPT's desktop client sits on everyone's machine. Equipping ChatGPT with LibreOffice effectively democratizes a tool that used to be a programmer's privilege. Sales teams can have Codex generate quotes, operations can drop in CSVs for pivot tables, and PMs can convert PRDs into Confluence format.

This is the reverse of a familiar pattern. The last decade moved tools from office suites (the BAT era) toward developer tooling (Git, IDEs, CI). Now OpenAI is pouring developer-tool capabilities back into the office world.

Deeper still: in Codex's current capability map, coding, documents, spreadsheets, and presentations form a continuous spectrum, not four separate tools. Local LibreOffice is what connects that spectrum.

Editorial Take

The real signal here isn't LibreOffice. It's that OpenAI is willing to add several hundred megabytes to an installer for a single capability.

In an era where everything moves to SaaS and clients keep getting thinner, going the opposite direction and stacking local binaries suggests the data-ownership calculus has changed in the agent era. Capabilities that can't leave the user's machine are the ones that get to call themselves "infrastructure."

Three predictions:

  1. Competitors will have to follow. Cursor's and Claude Code's biggest weakness in the enterprise market isn't coding ability; it's that they "can't open Excel." If Anthropic builds an enterprise Claude Code, it will probably bundle LibreOffice too, or acquire a headless Office company in the docparser or sheet2api mold.
  2. LibreOffice's standing will rise. For the past decade it's been the "open-source alternative," treated as a second-class citizen in enterprise IT. Now that OpenAI has picked it as the document runtime for the agent era, its commercial value and community resources will be repriced.
  3. The desktop agent moat isn't the model; it's client thickness. Claude and GPT-5 keep converging at the model layer. The real differentiator is who can call more local capabilities: the filesystem, Office, the browser, email, IM. OpenAI just turned "depth of client integration" into a new battleground.

One open question: is the several-hundred-MB bundle a stopgap or long-term infrastructure? If OpenAI eventually forks the UNO API into a stripped-down runtime, the way they forked Whisper, today's packaging could be replaced. Until then, LibreOffice's place inside agent clients is solid.


References:

  • Simon Willison, "Codex and LibreOffice," 2026-09-01, simonwillison.net
  • Hacker News discussion, item?id=49527396, 129 points / 72 comments
Source: https://simonwillison.net/2026/Sep/1/codex-libreoffice/