The editorial argues that repos like Codex-5.5-codex-instruct-5.5, FuckClaude, and chatgpt-register-sub2api contain no original code or models — they're thin wrappers that convert consumer ToS-bound subscriptions into metered programmatic API access. The naming conventions (sub2api as a genre) and keyword-stuffed titles are engineered specifically to intercept searches and juice velocity metrics.
The repo name 'Codex-5.5-codex-instruct-5.5' is itself the argument — it demonstrates the keyword-stuffing tactic by engineering a title designed to intercept every search query for OpenAI's coding model, exploiting the trending algorithm's reliance on velocity and search signals.
FuckClaude is a functional proxy that turns a Claude.ai web session into an API-shaped endpoint, directly embodying the ToS-arbitrage pattern the editorial critiques. Its high trending score despite being pure subscription-wrapping demonstrates how easily the ranking signal is captured.
chatgpt-register-sub2api automates ChatGPT account registration and pipes it into an OpenAI-compatible API surface — the flip side of the FuckClaude approach. It exemplifies the 'sub2api' naming genre that has become a recognizable pattern across Claude, Gemini, Grok, and Poe wrappers.
The editorial makes a structural argument: stars-per-unit-time works only when the ranked population is homogeneous in intent. Once a portion is optimizing directly for the metric — via keyword stuffing, coordinated timing around model releases, and anti-abuse-heuristic gaming — the signal breaks and Trending becomes a directory of exploits rather than genuine work.
The editorial highlights that Meituan's LongCat-2.0 — a genuinely open bilingual model with published benchmarks and real weights — failed to crack the top 20 the same week the proxy repos dominated. This inversion (wrappers above actual model releases) is presented as evidence that Trending no longer surfaces the work that matters.
The LongCat team shipped a substantive open bilingual model with published benchmarks, representing the kind of legitimate open-source release Trending was designed to surface. Its relatively modest score of 221 against gamed repos scoring 1,321 illustrates the discovery gap.
science-skills represents another category of substantive content competing for trending visibility against the proxy-wrapper genre. Its presence at 182 alongside FuckClaude at 320 shows how mixed the trending page has become between real work and metric-optimized exploits.
GitHub Trending this week surfaced three repositories that, taken together, describe the current state of open-source discovery better than any post-mortem could. The top of the list, at a score of 1,321, is `yynxxxxx/Codex-5.5-codex-instruct-5.5` — a repo whose name is engineered to intercept every search anyone has ever typed for OpenAI's coding model. At 320 there is `LinXiaoTao/FuckClaude`, whose title is not a joke about vibes but a functional description: it is a proxy that turns a Claude.ai web session into an API-shaped endpoint. At 178, `akihitohyh/chatgpt-register-sub2api` does the same trick from the other direction — automated ChatGPT account registration piped into an OpenAI-compatible API surface.
None of these repositories contains a model. None contains original research. All three are wrappers that convert a consumer subscription — one you agreed to a ToS for, and one the provider priced assuming a human on the other end — into metered, programmatic access. The `sub2api` naming convention is not accidental; it's a whole genre now. Search GitHub and you'll find a dozen variants targeting Claude, Gemini, Grok, and Poe, each one a thin Node or Python server that logs in, refreshes cookies, and re-emits `chat/completions`-shaped JSON.
Meanwhile, in the same week, Meituan's LongCat team quietly pushed weights for LongCat-2.0, a legitimately open bilingual model with published benchmarks. It is not in the top 20.
GitHub Trending has always been a velocity metric — stars per unit time, weighted by recency, with some anti-abuse heuristics bolted on. That works when the population of things being ranked is roughly homogeneous in intent. It breaks the moment a portion of the population is optimizing directly for the metric. And these repos are optimizing directly for the metric. The naming (`Codex-5.5-codex-instruct-5.5` is a keyword stuff, not a title), the timing (published in the days after OpenAI's Codex refresh), and the star patterns (bursts from newly-created accounts) are the same playbook you see on npm typosquats and Chrome Web Store clone extensions.
The deeper problem is that star velocity has no notion of *provenance*. It cannot distinguish a research lab dropping weights from a proxy that lets a Telegram channel resell someone else's API quota. To the ranking function they look identical: many humans, in a short window, clicking a star. If anything the proxies win, because their value proposition — free API access to a paid model — creates exactly the kind of viral loop the algorithm rewards. LongCat-2.0 has to earn its stars from ML engineers who read papers. `FuckClaude` earns its stars from every developer who wanted to script something against Claude and didn't want to pay for it.
There is also a legal geometry here worth naming out loud. Every one of these `sub2api` projects is, functionally, a scraper of a paid service being redistributed as a free one — and the repo owners know it, which is why the naming leans hostile rather than neutral. `FuckClaude` is not naming itself that way for edgelord points; it's naming itself that way because a repo called `claude-api` would get a DMCA within a week. The hostility is camouflage. It also happens to be excellent SEO.
Community reaction has been split in a predictable way. On Hacker News threads about similar projects, one camp treats them as a tax on providers who charge business prices for consumer subscriptions — the same argument that gets deployed for yt-dlp. The other camp points out, correctly, that these are the exact tools that get used to farm accounts, spin up phishing infrastructure, and quietly launder API calls through stolen credit cards until Anthropic or OpenAI's fraud team notices. Both camps are right. That's why the projects are popular. That's also why they shouldn't be the top of a discovery feed for developers looking to *learn what's happening in open source*.
The practical implication, if you rely on trending as a scouting signal, is that the top of GitHub Trending is now closer in character to app-store search results than to Hacker News: heavily optimized, systematically gamed, and biased toward whatever creates the tightest viral loop. That doesn't mean the signal is worthless, but it means you need a second filter. A few that work in practice:
First, weight by repo age and commit history, not by star count. A repo that appeared four days ago and has 1,300 stars and one contributor is a marketing artifact. A repo with two years of commits and 400 stars this week is a project that finally got noticed. Trending doesn't distinguish these; you have to.
Second, treat any repo whose name is aggressively SEO-shaped — model names, version numbers stuffed into the slug, or hostile prefixes — as guilty until proven otherwise. The naming is the tell. Real projects almost never need to name themselves `Codex-5.5-codex-instruct-5.5`. They can afford one word.
Third, cross-reference against actual release channels: Hugging Face for weights, arXiv for research, the release notes of the tool the repo claims to be about. If a `Codex 5.5` repo is trending and OpenAI hasn't announced Codex 5.5, the repo is not what it claims to be. This sounds obvious. It is not being applied.
For teams building on top of these APIs commercially, there's a second-order effect worth planning for. Every successful `sub2api` project shortens the timeline for the provider to tighten anti-automation on the consumer product, which usually means more captchas, more device fingerprinting, and more aggressive session invalidation for everybody — including the human users the product was designed for. Anthropic has already been visibly tightening web-session behavior; OpenAI's login flow got measurably worse in Q2. The proxies are the reason. If your workflow depends on stable browser access to a chat product, expect it to keep degrading.
GitHub is not going to fix trending, because trending is a growth surface, and the current mix — model announcements, viral CLIs, and grey-market API bridges — is producing more sessions, not fewer. The right move for practitioners is to stop treating the leaderboard as an editorial product and start treating it as what it is: a raw feed with a lot of noise, one that rewards the same behaviors that broke every other ranked discovery system on the internet. The irony worth holding onto is that in a week when a genuinely open bilingual model shipped, three wrappers around closed models outranked it. The wrappers are downstream of a discovery function that cannot tell them apart. Until the function changes, you have to.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.