Codeberg published a blog post explaining why it added a terms-of-service rule against projects that mostly consist of code written by generative AI tools such as Claude or Codex. Codeberg is a European nonprofit forge built around free and open source software, not a general-purpose GitHub replacement, and the post frames the change as both practical and cultural. The practical case is that LLM-assisted development makes it cheap to create large low-traffic projects that burn CI, storage, and moderation capacity. The cultural case is that Codeberg wants to host collaborative human-centered free software, not become a dumping ground for machine-generated ghost projects or a backstop for AI companies that trained on the commons.
A lot of the conversation settled on that identity point. People who liked the move saw it less as a universal claim about software engineering and more as a community saying what it is for. In that reading, Codeberg is leaning into a niche. If GitHub is the default forge, Codeberg wants to be the place where a repo still signals human stewardship and some resistance to AI slop. Several commenters said that kind of sorting may be the whole value. The governance angle helped this camp too. The rule came from member voting, and supporters treated that as a valid outcome for a nonprofit association rather than a betrayal of platform neutrality.
The strongest criticism was not that Codeberg lacks the right to do this. It was that the line is too fuzzy and the justifications are scattered. The actual terms-of-service text bans projects that mostly consist of AI-written code and cites copyright uncertainty and harmful code risk. The blog post emphasizes different concerns like community erosion, resource waste, and anti-extraction ethics. That mismatch made people worry about selective enforcement. Once you say "mostly AI-written" in an era of autocomplete, agents, and human curation across hundreds of prompts and commits, the rule stops sounding objective and starts sounding discretionary. For skeptics, that means the real cost is governance risk. Even teams that barely use LLMs may avoid Codeberg simply to avoid future arguments over proving what is or is not "vibecoded."
The comments also converged on a narrower reading of what Codeberg is probably aiming at. Despite the heated rhetoric around the policy, several people pointed out that Codeberg is not banning all AI assistance. The target is bulk generated projects with little evidence of real maintenance discipline, not every maintainer who uses Claude for packaging, debugging, or boilerplate. That softer interpretation made the policy seem more workable to supporters, but it also exposed the core tension. If the real problem is spammy low-quality projects consuming shared resources, a direct quality or pricing policy would be cleaner. By choosing LLM use as the proxy, Codeberg gets a stronger cultural signal and a murkier enforcement problem.
If you host code on a community-run forge, treat it less like neutral infrastructure and more like a values-driven club with real moderation risk. If your workflow depends on heavy AI assistance, expect some open source communities to make that an explicit compatibility question, not just a tooling preference.
Mixed but polarized. Many approved of Codeberg drawing a clear anti-LLM boundary for a community-run forge, while an equally vocal group saw the policy as vague, moralizing, and likely to invite arbitrary enforcement more than it solves resource abuse.
Key insights
01
TOS text and blog post do not match
The actual terms-of-service change bans projects that mostly consist of AI-written code on copyright and harmful-code grounds, while the blog post sells the move on community health, resource use, and anti-extraction values. That gap changes how you read the policy. It suggests the public rationale is broader and more emotional than the enforceable rule, which is exactly why people started asking where the real boundary sits.
Read the policy text, not just the announcement post, before betting your project on a host. If the legal rule and the cultural explanation diverge, assume moderators will have wide discretion later.
Several commenters read the move as an attempt to make a Codeberg repository mean something again. The value is not just file hosting. It is the expectation that a project was curated by humans who care, rather than mass-produced to satisfy a prompt. That turns the forge itself into a quality signal, much like older communities used admission barriers to keep the commons usable.
Choose hosting based on the signal it sends to contributors and users, not just features. A smaller forge can become more valuable if its moderation policy makes the average repo more trustworthy.
The sharpest practical reading is that cheap code generation breaks the economics of donation-funded hosting. If one-person ghost projects can now produce enterprise-sized repos, CI pipelines, and release artifacts, shared infrastructure gets swamped fast. Seen that way, the anti-LLM rule is less a philosophical judgment than a rough throttle on unbounded volume.
If you run shared developer infrastructure, price and quota assumptions built for human output may already be obsolete. Revisit storage, CI, and abuse controls before AI-assisted usage forces you into blunt policy moves.
A recurring concern was not hostility to Codeberg's values but the cost of relying on a host whose boundaries can narrow over time. People worried less about today's rule than about tomorrow's reinterpretation, especially when the post admits there is no easy line and enforcement will be judgment based. That makes Codeberg feel less like neutral plumbing and more like a club whose norms can shift under your feet.
For business-critical or long-lived projects, ask whether your host behaves like infrastructure or like a mission-driven community. If policy drift would be expensive for you, self-host or keep migration paths warm.
Git hosting and project discovery are separate choices
A useful side point was that Codeberg can work well as a forge even if it cannot replace GitHub's network effects. GitHub still provides the social graph, stars, follows, and default contributor accounts that make projects discoverable. Codeberg offers a place to host code under different norms, but it does not automatically bring the same audience.
Do not confuse repository hosting with ecosystem reach. If discoverability and casual drive-by contributions matter, you may need GitHub even if you mirror or primary-host elsewhere.
One maintainer argued that Claude materially improved packaging, CI, issue triage, deployment, and upstreaming fixes, cutting backlog and making contribution easier rather than more extractive. That pushes back on the idea that AI use mainly creates slop. For at least some maintainers, it increases participation in free software and unlocks work that otherwise would not get done.
Do not assume AI-assisted contributors are net negative by default. If you maintain a project, measure whether tooling is reducing backlog and improving submissions before copying a blanket policy.
A strong dissent said the problem worth policing is low-quality spam, not LLM use itself. Experienced developers can use models heavily and still ship careful software, while bad manual code has always existed. Targeting the proxy instead of the outcome may just push people to hide their tooling and leave moderators stuck with the same quality problem.
If your real pain is junk submissions or runaway CI use, write policies and tooling against those behaviors directly. Proxy bans feel decisive but often create evasion without fixing the operational burden.
Resource limits would be cleaner than moral framing
One commenter argued Codeberg is backing into ethics and copyright arguments to justify a simpler reality: it built for a world where code volume was naturally constrained, and that world is gone. Pricing, quotas, or stricter project limits would address the infrastructure problem more honestly than trying to sort worthy from unworthy code by origin story.
When capacity is the issue, solve capacity first. Clear quotas and paid overages are easier to defend and administer than intent-based moderation rules about how code was produced.
blog.codeberg.org