Two million downloads. That is what The Document Foundation reported for LibreOffice 26.8, the release whose headline characteristic is the absence of generative AI. The first million arrived within a week of the announcement. For an office suite that has been shipping competently and quietly for years, the fastest growth came not from adding a capability but from publicly declining one.
I spend most of my time looking at how agent systems are wired together, where inference happens, what crosses a process boundary, and who holds the keys. From that vantage point, the LibreOffice story is less about sentiment toward AI and more about a specification that users can actually check.
Absence stated as an architectural property
Read the claim carefully. LibreOffice 26.8 contains no generative AI features, does not transmit documents to remote servers, and requires no network access to function. Those are three separate statements, and the last two are the interesting ones.
“We respect your privacy” is a policy. “The application functions with the network interface down” is a property you can test with a firewall rule and a packet capture. The difference matters enormously in agent architecture, because almost every assurance offered by assistant features in productivity software today lives on the policy side of that line. Your document goes somewhere. A model you cannot inspect processes it. A retention promise you cannot verify governs what happens next. The user is asked to trust a contract rather than a boundary.
What LibreOffice shipped is closer to a verifiable invariant. Whether or not you care about AI, that is a different class of guarantee, and I suspect a meaningful share of those two million downloads came from people who intuited the distinction without having the vocabulary for it.
The trust boundary is the product
Agent designers tend to treat the trust boundary as an implementation detail, something to document in a security appendix after the capability demos are done. The download numbers suggest users experience it as the primary feature.
Consider what an in-document assistant actually requires. It needs read access to the full document, because context is what makes the output useful. It usually needs write access, because suggestions that cannot be applied are not worth much. It needs an egress path to wherever inference runs. And it needs to be resident, watching, because an assistant that only works when summoned loses most of its perceived value.
Stack those requirements and you have described a component with total visibility into a user’s work and a channel out of the machine. That is a reasonable thing to build. It is not a reasonable thing to build by default, without consent, in software used by tens of millions of people. The Document Foundation did not reject AI outright. Its position was that certain principles have to be satisfied first, given who depends on this software and for what.
Collabora’s opt-in path is the honest middle
The nuance here is worth sitting with. Collabora Office 26.04 introduced AI as an opt-in feature, off by default. Same broad codebase lineage, different product decision, and both can be defensible at once.
Off-by-default is the architecturally honest configuration. It preserves the invariant for every user who does nothing, and it moves the trust boundary only when someone deliberately moves it. Compare this to the pattern that has become common elsewhere, where assistant features arrive enabled, with a settings panel somewhere three menus deep, and the burden falls on the user to discover and revoke. The defaults encode the real position, whatever the blog post says.
For those building agents into existing tools, this is the design question that actually matters. Not whether to offer agentic capability, but whether the system still has a clearly stated, checkable behavior when the user has not opted in.
Reading the numbers with appropriate caution
I will add a methodological note, because I would want one from someone else. Commenters on Hacker News picked at the download figures, pointing out that the release-week numbers look different when viewed by day rather than by week. Release spikes are noisy. Download counts are a weak proxy for retained users, and a strong proxy for attention.
So treat two million as evidence of demand signal rather than a measurement of adoption. That signal is still real, and still useful. It says a substantial population will go out of its way for software that makes a legible promise about where their data stays.
What this means for agent work
The lesson I take is not that AI features repel users. It is that unverifiable claims do, and that the market for checkable guarantees is larger than most product teams assume.
Local inference, explicit consent gates, documented egress, and an off-by-default posture are not concessions to skeptics. They are the features. An agent whose boundaries a user can reason about will beat a more capable one whose boundaries are a matter of faith, at least among the people who read release notes. Two million of them, apparently, read this one.
🕒 Published: