Codeberg Bans Vibe-Coded Projects: What Actually Changed
Codeberg's members voted to prohibit vibe-coded repos over hosting costs, not code quality. Here's the actual vote, the reasoning, and the self-hosting community's reaction.
Ask most people what's wrong with vibe-coded software and you'll get an answer about code quality — hallucinated logic, missing edge cases, security holes nobody reviewed. Codeberg's members just voted to prohibit vibe-coded projects on their platform, and code quality is barely in the reasoning. Their actual argument is about who pays for hosting a mountain of software nobody's using.
What actually got voted on
Codeberg e.V.'s annual member assembly runs two-week async votes on proposals raised at the live meeting, and this July two of them concerned AI directly. Both passed. The first was a straightforward reaffirmation: Codeberg will not use hosted code or user data to train any LLM, full stop. The second was the contentious one — a change to Codeberg's Terms of Use prohibiting "vibe-coded projects" — and it passed 358 to 144, with roughly 50% of active members turning out to vote.
The reasoning laid out in Codeberg's own writeup is worth taking at face value rather than assuming it's just anti-AI sentiment. Their core complaint is resource asymmetry: they describe seeing single-developer, effectively-userless projects generating as much CI/CD and storage load as some of the largest genuinely-community-maintained projects on the platform — heavy test suites, frequent large release binaries, broad platform support that outstrips any plausible user base. Layered on top of that is a hardware cost problem that's out of their control: a type of drive Codeberg says cost roughly €700 a few years ago now runs about €3,700, a squeeze they attribute partly to AI-driven demand across the storage and memory market.
The exact rule, and the fight over "mostly"
The change didn't arrive as vague policy language — it went through Codeberg's normal contribution process as an actual pull request, #1253 on the Codeberg/org repository, titled "Proposal Assembly 2026: ToU extension to prohibit LLM-extrusions." It was opened by a contributor going by gedankenstuecke, with sign-off from Codeberg board member Gusted, and merged only after the member vote passed at the 2026 general assembly.
The operative line added to Codeberg's Terms of Use targets any project that would count as containing "projects that mostly consist of code written by" generative AI tools — and the ToU doesn't leave the target vague, naming Claude and OpenAI Codex explicitly as example tools covered by the rule.
That single word — "mostly" — is where nearly all of the public argument has actually landed, more than on the ban's existence at all. Codeberg's own contributors debated the wording in the pull request itself before it ever went to a vote, and once the results went public the same question spread to Hacker News and Reddit: mostly measured how, exactly? Commenters raised the obvious edge cases — does heavy use of IDE autocomplete count as "mostly AI-written"? What about IntelliJ's automated refactoring tools, which also generate code a human didn't type by hand? Nobody, including Codeberg, has published an objective threshold. One commenter defending the wording argued it isn't meant to be forensically precise — it's meant to capture a pattern moderators say they're actually seeing: reviewers rubber-stamping huge LLM-authored pull requests without meaningfully reading them, not any specific percentage of AI-assisted keystrokes.
The actual scope of the ban
This isn't a purge, and Codeberg was explicit about that. Per their own early guidelines, a project is unlikely to be touched at all if it has an active community actually maintaining it, meaningful history that predates heavy LLM use, or if a maintainer simply received some LLM-generated contributions from outside contributors without the project itself being LLM-driven.
| Likely fine | Likely tolerated | Likely unwelcome |
|---|---|---|
| Actively maintained project with a real community | Small side projects/experiments with low resource use | Projects created autonomously by LLM agents |
| Significant pre-LLM project history | Narrow custom scripts unlikely to attract a community anyway | Projects heavily tied to the LLM ecosystem itself (e.g. LLM tooling to ease LLM usage) |
| Occasional LLM-assisted external contributions | — | Resource footprint clearly exceeding what the actual contributor base could produce by hand |
Codeberg also says outright that they won't be running automated scans to hunt down offending repos — enforcement is going to be case-by-case and human-judged, not a bot sweep. Their stated worry runs deeper than any individual repo anyway: they describe a "vicious cycle" where LLM-assisted contributions increase review burden on maintainers, trust between contributors erodes as people start accusing each other of using AI (sometimes wrongly), and the net effect is that collaborative projects become less appealing to work on than just vibe-coding a disposable personal tool instead.
Not everyone's convinced — including inside Codeberg
The 358–144 result is a clear majority, but it's not unanimous, and some of the pushback carries real weight. Armin Ronacher — creator of the Flask web framework and a co-founder at AI agent company Earendil — publicly called the decision "a very bad move" and urged Codeberg's leadership to reconsider. Coming from someone with that much open-source credibility, it's a meaningfully different objection than the general internet grumbling the vote also attracted.
More pointed criticism came from inside Codeberg's own membership about the process itself, not just the outcome. One member's account of the live assembly meeting, shared on Hacker News, describes asking a question about the server-load justification and getting an answer about crawler-bot scraping rather than anything about AI-generated code specifically — and characterizes the live meeting as having arrived with the text already finalized, functioning more as a Q&A than an actual discussion. The same account questions how representative the vote really was, noting uncertainty over how many of Codeberg's members engaged with the substance versus simply responding to what they characterize as an email amounting to a yes/no prompt on banning AI.
It's also worth knowing the AI vote didn't happen in isolation. The same assembly separately voted to ban cryptocurrency projects from Codeberg too, explicitly citing SourceHut's earlier 2023 decision to do the same as precedent. Read together, the two bans paint a consistent picture: a resource-constrained nonprofit forge drawing a harder line against categories of projects it associates with high infrastructure cost and low community payoff — not a rule invented specifically to target AI as a technology. That framing hasn't fully landed with critics, though; some of the same skepticism aimed at the AI ban showed up on the crypto-ban thread too, with commenters questioning why resource-intensiveness alone should make a category of source code unwelcome at all.
This is where it gets more interesting than a policy story. The self-hosting community's own reaction wasn't really outrage at Codeberg — it was a newsletter writer turning the mirror on himself.
Ethan Sholly, who runs the widely-read Self-Host Weekly newsletter, spent a week vibe-coding two personal homelab tools with a $20 Claude Pro subscription: a Kindle/Audiobookshelf progress-sync tool and a YouTube-to-podcast-RSS transcoder. He described the process as genuinely satisfying — then decided not to publish either one, citing uncertainty about which upstream projects or licenses his AI-generated code should credit, and, in his own words, not having thick enough skin for "the wrath of this community." That last line was a direct callback: the prior week's newsletter had already been mockingly rebranded "Slop-Host Weekly" by readers in the feedback.
The satire didn't stop at a nickname. In direct response to Codeberg's policy change, a newsletter reader pointed the community toward shitcode.org, a site that catalogs deliberately terrible AI-generated repositories as a kind of running joke/warning — and Sholly found and shared one specific example from it in a subsequent newsletter. And the scope of what's being vibe-coded keeps climbing: full operating systems built primarily through AI-assisted prompting are now a real, shipping category, not a hypothetical.
What's out of scope here
This isn't a story about vibe coding being banned everywhere, or about a consensus forming that AI-assisted code is categorically bad. Codeberg's own guidelines carve out plenty of room for incidental or well-integrated LLM use, and nothing in their vote touches GitHub, GitLab, or any other forge — this is one platform's specific resource-economics decision, not an industry-wide verdict. It's also not really a story about code quality at all, despite how the meme culture around it reads; Codeberg's stated objections are almost entirely about hosting costs and community erosion, not about whether the generated code works.