Drag to orbit · scroll to zoom · click any module — launch the live 3D tour
I joined the GNU Hurd project on Savannah back in 2021. The learning curve was — and still is — brutal: a multiserver microkernel, GNU Mach, translators, capability-passing ports, decades of mailing-list history. I never shipped a patch(It's hard for me(GitHub generation) to use email to do the PR and it my problem of course).
That gap is the reason this repository exists. sites/hurd is an interactive three.js explainer of the Hurd's microkernel core, its components, and its boot sequence — the field guide I wish I'd had in 2021, built so the next person doesn't bounce off the learning curve the way I did.
It's also a small experiment with a bigger question underneath it, explained below.
My working idea was simple: modern AI coding agents could make a real dent in the Hurd's backlog — triaging decades-old bugs, mapping unfamiliar subsystems, drafting patches for a project that is chronically short on contributor-hours.
So I emailed a few other Hurd contributors to talk it through, prompted by the project's own 2026 Q1 news post. That post already answers the question, and not the way I expected:
Brent W. Baccala worked with a Claude AI bot to debug x86_64 SMP issues. The bot did not contribute any code — it just found some incorrect code, which a human contributor then fixed. It also got some things wrong.
The Hurd is happy to let an AI read — to scan, to flag, to point at a suspicious spin_lock or a stale comment. It draws a hard line at letting an AI write the fix. A human has to understand the code well enough to author the patch and take responsibility for it.
That answer is what sent me down a rabbit hole: what does "free software" actually mean, once you compare it to how "open" software gets built today?
The Hurd's line isn't about code quality — Claude found real bugs. It's about authorship and accountability, which the Free Software Foundation has always tied to the four freedoms, not to a license file:
The four freedoms, for reference
| # | Freedom | In practice |
|---|---|---|
| 0 | Run the program, for any purpose | No EULA telling you what you're allowed to use it for |
| 1 | Study how it works, and change it | Source you can actually read, on a system you control |
| 2 | Redistribute copies | Help your neighbor, no gatekeeper in between |
| 3 | Distribute your modified versions | The improvement goes back to the commons, not to one company |
Compare that to how a lot of "open" AI work happens in 2026: weights you can download but training data you'll never see, "open" models built by companies whose product is the model, code-review bots with no accountable human behind the diff. It's open in the marketing sense, not in the Freedom 1 sense — you can't actually study or verify most of what made the thing behave the way it does.
The Hurd's rule — AI may look, a human must write and sign off — is a small, deliberate act of keeping authorship where the four freedoms put it. That's the real distinction I was missing when I first framed this as "AI agents will help the Hurd." The interesting part was never whether AI can help. It's who stays accountable for the result.
- GNU Hurd — official project page
- GNU Hurd — Wikipedia
- GNU Hurd Savannah project & member list
- GNU Hurd news, 2026 Q1
- What is free software? — GNU