My agent did it? I didn’t do it.

I don't see the problem?

Bundling LibreOffice with a chatbot app feels like peak software bloat. My hard drive just let out an audible sigh.

(I made a similar comment in a different post by Simon Willison yesterday about GPT Work)

I am unsure the direction that AI assisted document creation work will go in. I am opinionated and building something in the space. In this reply I discuss my view + what I have been building.

My opinion is: We now have AI, this new amazing flexible tool. We can throw it at any problem. But in terms of document creation we have a choice between old fashioned siloed document formats (.pptx, .pdf, .xlsx, .docx, etc.) and a freshly rendered - slightly randomly styled - html site with some basic javascript within it. I think these combinations are odd and do not maximise the powers of AI.

I've been building something called SmallDocs [0][1] which is a new take on work documents. It's all Markdown, but it always renders (100% privately) on the SmallDocs site. This means the SmallDocs renderer can convert recognised Markdown combinations into powerful traditional document formats [2]. For example, we convert a ```cells block into an in-line spreadsheet you can export to Excel.

This gives an agent the ability to express itself with a great deal of depth and flexibility. For example, your agent can produce a single data analysis "SmallDoc" with text, charts, spreadsheets, etc. [3] I find that combining formats quickens the speed I can understand a topic.

SmallDocs also have standardised and pretty boring styling. This means you spend less time understanding the layout of a flashy HTML page and more time on the content.

I also believe a pre-determined rendering engine has a lot of advantages over a freslhy built HTML artifact. It allows you to build in lot's of useful defaults that the agent doesn't have to code from scratch. These can be small things, for example, every time your agent renders a Markdown table in SmallDocs you can instantly copy it as a CSV or a PNG. And these can be bigger things, such as a rich slides DSL with solid styling defaults, which saves your agents a lot of tokens and results in consistently attractive HTML slides [4]. (Slides can also be automatically exported to .pdf and .pptx.)

I've been at it for a few months, and am still working on the project. I am looking for feedback and open source collaborators. Thanks for reading!

[0] https://smalldocs.org

[1] https://github.com/espressoplease/smalldocs

[2] https://smalldocs.org/blogs/what-is-a-smalldoc

[3] https://smalldocs.org/s/46fWiAyF77T9c7tqYwAQqp#k=j78ZEOOrgo8...

[4] https://smalldocs.org/s/QrmeWMsULGS871nXQYOacB#k=ubOHH9QtTWp...

(And something totally different, your agent can also use SmallDocs to walk you through code files: https://smalldocs.org/s/JsI_-tWSH0UpSd_YXxVLPm#k=G5jLCYA4JUX...)

And wait till you find out about @oai/walnut

I still don’t understand why with all their cash these companies dont redo their apps in Rust or native. They are all so heavy and dog slow. There are already some very fast cross platform (wasm too!) libraries that can do what poppler does as well as parse and produce various office file formats. Because 2gb is crazy town, these things could be around 100mb, just throw some tokens at the problem.

Kind of. Eventually it gets too slow even for the OpenAI engineers using it, and then they need to fix it.

True, I feel like the entire HN community is to blame for this, particularly pg and dang.

I have a feeling you should learn about Poe's law: https://en.wikipedia.org/wiki/Poe's_law

one does not just "throw tokens" at a problem, Ai is not that capable, and we are talking about MS Office file format, likely a NP hard problem

On top of everyone else’s great points, if you make your own “fork” of something, then you have to consistently maintain it. Which can be done with tokens, but an extra cost with something that’s not as user tested.

Especially when there are three different versions, and each has a different team, and they all can work together at the same time. Crazy amount of tech supporting it that breaks a humans context let alone a models context.

A full implementation/rewrite of something like word, which is pretty complex and the quantity of edge cases is insane, in a relatively short amount of time, months, is very hard.

If they succeed with AI it will be huge: it's basically not possible to do it today with humans, no matter how many people you put on the job.

I’m not sure about you but I would not want to adopt and ship an overnight “word” written in rust

> Whole lotta words to say that I don't know whether this is a GPL violation,

I don't need a lot of words to say it can't be a GPL violation since Libreoffice is distributed under the MPL v2.0 license.

https://www.libreoffice.org/licenses/

https://www.mozilla.org/en-US/MPL/license-policy/

I finally did find the license in

  ~/.cache/codex-runtimes/codex-primary-runtime/dependencies/native/libreoffice-headless/libreoffice/LibreOfficeDev.app/Contents/Resources/LICENSE

This is from a recent update, and I'm on macOS. I had to re-download it, not just update from within the app.

This new version has tabs for ChatGPT and Codex. The ChatGPT tab seems way less powerful; Codex is basically a coding agent with a UI and browser built in.

Pretty sure this is the old Codex app with the ChatGPT tab bolted on. It's great in some ways, but needs work in others.

For those that are interested, this inspired me to create a tool just for exporting images and PDFs of Word, Excel, and PowerPoint files. Check out the repo below, or if you have uv installed just run `uvx office-export skill install` or `uvx office-export --help` to check it out.

It uses Office for Windows, so it is Windows only. I don't currently have access to a Mac machine, so I can't build in Mac support yet.

https://github.com/pseudosavant/office-export

Thanks for sharing, now im wondering if they did in fact grant my wish since I messed with this in January. Im lucky my job does not require me to make any pptx presentations.

I think they meant to view documents.

This is one that's already true and "everyone" hasn't figured it out yet. But you don't really even need Office as a viewer, PDF is fine (for printable/portable fixed-layout) or just HTML. "Office" is a product that doesn't need to exist anymore, at least for the core apps. For day-to-day office work, you probably shouldn't be using those legacy apps, you should be reviewing portable document formats and asking an agent to make the revisions that you want.

> For Microsoft, this shift means products like M365 will act as the foundational layer where AI agents operate. Users will still need places to store documents, manage projects, search files, and track workflows

They bet on the online office suite (M365).

... and without the license, crucially!

> The software will suck and will not really innovate, but it will make a ton of money

Microsoft’s goal isn’t to provide high quality software that outperforms other software, their goal is to provide an obvious and reasonably proximate answer to clueless execs trying to check things off of a procurement list so that existing contracts can snowball and competition is excluded.

We, collectively as a global species, really lost out when MS wasn’t divvied up in their anti-trust trial. Office as its own corporate entity would be in a radically different place right now, as would their enterprise customers.

Yes, although it's buried among a huge number of other shortcuts now, and not particularly easy to search for.

I don't think that's anywhere near equivalent. That's for reading and writing Microsoft Office docs. It has no office functionality. LibreOffice has a full API, for example to run calculations on spreadsheets [1].

The goal of bringing in dependencies is so you don't have to duplicate the functionality, in this case, crafted over decades of effort.

[1] https://wiki.documentfoundation.org/Documentation/SDKGuide/C...

Cool project! I'm playing around with some similar stuff, but via uv/Python. Impressive you aren't leveraging any other existing libraries for OpenXML parsing, editing, or rendering. Any especially tricky gotchas you've run into for this tool?

How do you mean?

I trust OpenAI to have installed apps that at least work well enough for their intended purposes.

If you've ever actually gone down this route you will try exactly this realize how much of a cluster f these products are and just use libre.

I did the exact same thing in my AI system. Spent an ENOURMOUS amount of time trying everything I could and in the end libre was the only thing even somewhat functional.

Easily 2 months of my life lost to these god forsaken systems. APIs and CLIs sound reasonable until you actually attempt anything in this space.

Computer use is exactly the same thing. Go try this: go to a google doc and dev tools check out the dom and now do the same for a word doc. Eye gouging pain of trying to get anything to work.

Computer use is slow and can take control of the application focus. I’ll be doing things while Claude runs in the background. Suddenly it decides to open the app and try to do something + screenshot it.

On a Windows installation I see this:

    PS C:\Users\user> ls C:\Users\user\.cache\codex-runtimes\codex-primary-runtime\dependencies\native\
    Mode                 LastWriteTime         Length Name
    ----                 -------------         ------ ----
    d-----         8/27/2026  10:09 AM                git
    d-----         6/16/2026   9:18 PM                jxrlib
    d-----         8/26/2026   7:37 PM                libheif
    d-----         6/17/2026   5:32 AM                poppler
    d-----         8/27/2026  10:09 AM                powershell

So doesn't seem bundled with all installs at least.

Notably: this did not exist after I installed the app, it existed after I ran the app and there was the long startup screen. Presumably, and I'm not sure I'm interested enough to redo it all and find out, that long startup screen involves the default skills plugins setting up their environment and one of those pulls it as a general app requirement.

I also saw this present after first run on macOS but not Windows.

I have had the Codex/ChatGPT app installed for a few months, and can confirm that there's a 422.9 MB file named LibreOfficeDev.app in the above path.

More notably, I don't use Computer Use or any of the fancy features, just Codex for agentic coding.

Do you have any tips for running Codex on the Macbook Neo? Have you changed any setting? On mine, it stutters on every single chat, and I don't even have any plugins installed.

> Mac Neo can be had for $600 now

Maybe a used one?

https://www.apple.com/macbook-neo/ has it for ~$700

They're also great at powershell, you just need to tell it to do so. sh is neat but it's 55 years old, there are newer and better shell environment concepts out there and the robots can handle them well too, try it some time!

Ironically, I feel like LLMs have really unlocked Windows work since they all know how to use Powershell and all the weird Windows commands/hacks that you otherwise either wouldn't even know about, or would have to Google every time.

I used to hate working in Windows, but now it feels roughly comparable to Mac (though at the risk of inciting an OS war, I do feel like Linux still beats both out...)

How do you work on Windows in the age of LLMs?

That’s easy: you activate WSL and then ask the LLM how to get networking and stuff working since they changed it with WSL2. You need a config file flag to setup network mirroring now.

It does provide a pretty good COM interface.

And libreoffice has some really useful and easy to use tools for document manipulation and creation hiding in the background.