Simon Willison discovered earlier this month, while poking around in ~/.cache with OmniDiskSweeper, that OpenAI's Codex desktop app (since rebranded to ChatGPT) had stuffed a 1.7GB folder called codex-primary-runtime into its cache directory. Inside: a full LibreOffice, a full Python, a full Node.js, plus Poppler and git. The documents subdirectory contains skills that teach Codex how to call these binaries.

Why a desktop AI assistant needs LibreOffice

PDF extraction, docx parsing, table reading — the model cannot do these itself, real tools must execute them. By bundling LibreOffice, Codex can convert PDF to docx, docx to PDF, or run in headless mode for batch conversion. Python and Node.js provide the runtime substrate that lets skills invoke arbitrary scripts. This is a micro AI workstation, not a chat client.

1.7GB versus 10GB: two different paths

But the real counterpoint is Anthropic's Claude Desktop. GitHub issue #22543 has been documenting this for months: every Cowork session writes a 10GB rootfs image to ~/Library/Application Support/Claude/vm_bundles/claudevm.bundle/rootfs.img, never cleaned up, regrows the day after deletion. On an 8GB Mac, idle CPU sits at 24%, climbing to 55% within minutes, swap steadily rising. The temporary workaround is violent: rm -rf three cache directories, task speed improves 75% — but only temporarily, the underlying memory leak is unfixed.

OpenAI's 1.7GB and Anthropic's 10GB differ in packaging philosophy. Codex follows the "bundle all native binaries" route: small footprint, fast startup, depends on system libraries. Claude follows the "embed a full Linux VM" route: stable functional isolation, but disk and memory overhead are an order of magnitude higher. Both paths point to the same conclusion: if an Agent is to do real work locally, model weights alone are insufficient — it needs a complete traditional toolchain.

The Agent paradigm is reverse 1990s-ifying

This is the paradigm inflection behind Agent products. The earlier Copilot generation placed code generation in the cloud IDE, with only a lightweight plugin locally. The new Agent generation reads files locally, runs scripts locally, manipulates Office documents locally. Calling local tools directly is more realistic than stacking every capability into the cloud model. The local toolchain is unavoidable — when users hand the Agent real documents, real PDFs, real code repos, the Agent has to treat them as binary objects in the local environment.

The bill lands on user disk and memory. 1.7GB is not extreme — but codex-primary-runtime is a single app's cache; Claude Desktop's vm_bundles sits at a permanent 10GB. A developer running Codex, Claude Desktop, and Cursor concurrently consumes 30GB in local AI toolchains. Layer in Hugging Face caches, uv, npm, Python venvs, and a 512GB laptop fills fast. Willison's screenshot shows huggingface alone at 356GB.

The industry's next move is either to learn from Docker Desktop — image dedup and on-demand loading — or to learn from cloud IDEs — push the toolchain to remote containers, leave only a lightweight client locally. From 1.7GB to 10GB, we see the product decision; what we do not see is the Agent paradigm eating the desktop back into a 1990s-style heavy-client era where you "install the whole toolchain."

So before installing the next AI client, check what it intends to stuff onto your disk.