The ChatGPT/Codex app bundles a full copy of LibreOffice

Original Article ↗ Hacker News Discussion ↗

OpenAI’s ChatGPT/Codex desktop app has been found to bundle a large headless LibreOffice installation (and other heavy runtimes) to read, write, and render Office documents locally, prompting debate over software bloat, attack surface, and licensing hygiene. Commenters note that, despite the size cost and overlapping with tools users may already have installed, LibreOffice is one of the few robust, scriptable options that can reliably handle messy, legacy Office formats—highlighting both the grim reality of office file complexity and the limits of having AI “just build” its own stack. The thread also broadens into criticism of the new app’s UX and performance, and speculation that as AI agents increasingly manipulate documents directly, suites like Microsoft Office may drift toward being primarily viewers rather than authoring tools for humans.

What was discovered

  • The desktop ChatGPT/Codex app installs a ~400+ MB headless LibreOffice runtime into a cache directory, apparently for document handling (DOCX/XLSX/PPTX, PDFs, conversions, screenshots).
  • On some systems it appears after first run / long “initialization” rather than at installer time. On Windows, other native tools (git, poppler, etc.) are also downloaded into the same cache.

Why LibreOffice and what alternatives exist

  • Several commenters say they independently ended up bundling LibreOffice because:
    • Office/Google APIs and CLIs are fragile, fragmented, or hard to automate reliably.
    • Old/odd Excel files and complex documents are best handled by LibreOffice.
  • Alternatives mentioned: Microsoft 365/Google Workspace CLIs, Apache POI, pandoc, custom PowerPoint/Office exporters, and a .NET OpenXML backend (“walnut”).
  • Some note that libraries like POI can compute formulas but don’t replace a full office suite API.

Bloat, performance, and resource concerns

  • Many see bundling a full office suite as “peak bloat,” especially given:
    • Some users already have LibreOffice installed.
    • SSD wear, large footprints, and multiple heavyweight AI apps/VMs on localhost.
  • Others argue storage is cheap relative to tokens and dev time; a known, battle-tested dependency is reasonable.

Security, licensing, and provenance

  • Concern that bundling large third‑party binaries expands the attack surface and may include unpatched vulnerabilities.
  • Questions about license compliance; a LibreOffice MPL license file was found inside the cached bundle.
  • Unclear if this counts as “bundled with the installer” or “downloaded on demand,” but most think MPL obligations are likely met by including the license.

App behavior and UX

  • Mixed experiences with the new ChatGPT/Codex app:
    • Some report good performance and like the 3‑pane, coding‑agent-centric UI.
    • Others call it messy, slow, resource‑hungry, and confusing (ChatGPT vs Codex, Chat vs Work splits, missing/changed shortcuts).
    • Windows users in particular report hangs, high CPU usage, and difficulty accessing mobile chats.

Implications for Office and AI

  • Several see a trend where Office becomes mostly a viewer while AI agents generate and modify documents via open formats and non‑Microsoft tooling.
  • Some think this highlights both the complexity of Office formats and current limits of AI: it’s still easier to ship LibreOffice than to have an AI “reimplement Word.”