Codeberg does not want to ban developers from opening Claude next to their editor. The free-software forge is drawing a more mundane line: it does not want to pay to store, serve and moderate a mountain of generated software that nobody really intends to maintain.

That is much less dramatic than "Codeberg bans AI." It is also much more interesting.

On July 23, the association behind Codeberg published the result of two internal votes around LLMs. The first says hosted data will not be used to train its own models. The second prepares a change to the terms of use around heavily vibe-coded projects. That motion passed with 358 votes in favor, 144 against and 14 abstentions.

Put like that, you can already imagine a tiny moderation robot sniffing commits for traces of Claude. Codeberg explicitly says it does not want to do that.

The problem starts before the code

Codeberg's post begins with infrastructure. Crawlers that ingest software forges do not always perform a polite git clone. They also walk file pages, histories, filtered views and other web surfaces that consume compute and bandwidth.

For a giant commercial platform, that becomes another line on a bill. For an association providing a free forge to the FLOSS community, the question gets stranger: why should its members finance infrastructure for companies coming to ingest their work?

That is Codeberg's argument, and it should stay attributed to Codeberg. At least it points to an identifiable cost, rather than turning the whole thing into a culture war between "real developers" and people who discovered Cursor last Tuesday.

Then comes the second cost: the repository itself.

Generating a disposable tool has become extremely easy. Giving it a name, a README, an icon and pushing it to a forge is easy too. What has not become equally cheap is understanding that repository, answering its issues, reviewing contributions, checking licenses or simply deciding whether it is still alive six months later.

Code got cheaper to produce. Attention around code did not.

A public repository is not automatically a commons

This is probably the part of Codeberg's argument I want to poke at the most.

We have grown used to treating "visible source code" and "open source" as almost automatic value. Put the project on a public forge, add a license, done: contribution to the great communal pot.

Codeberg pushes back on that equation. A script generated to solve your Tuesday problem, never read again and maintained by nobody, may be perfectly useful. It is simply not necessarily a community project.

The distinction is healthy. It also predates LLMs by decades. GitHub has always contained graveyards of side projects typed entirely with human fingers. AI mostly changes the scale. We can now dig the graveyard with an excavator.

The scarce thing is no longer the source file. It is maintenance, discussion, context and the fact that somebody still answers when something breaks.

The rule is deliberately messier than the slogan

Codeberg is not planning a mass deletion of generated projects. Its initial guidance explicitly protects projects with an active community, older projects using LLMs occasionally, and probably plenty of small experiments that consume little infrastructure.

Projects produced autonomously by agents, maintained overwhelmingly through LLMs or disproportionate in storage and CI relative to the humans involved may, on the other hand, stop being welcome.

That is fuzzy.

I almost prefer it that way.

A perfectly automatic rule would force Codeberg to invent a magic percentage of human-written code, scan repositories, or pretend we can cleanly infer where every file came from. We cannot. Neither can they. So the organization is defining a community rule before pretending it has a detection rule.

There will be annoying edge cases and arguments. But the useful question may not be "how much code must a human type?" It is what we expect from a forge funded as a commons.

Hosting source code is no longer necessarily the expensive part. Deciding which code deserves continued community attention is starting to look like it.