The editorial argues that FuckClaude and sub2api are not pranks but productized versions of a months-old arbitrage: Claude Pro at $20/month is 10-100x cheaper than the API for the same underlying model, and the community has built a cottage industry to harvest that gap. It frames the coordinated trending across Chinese, Korean and Western repos as evidence that the subscription-vs-API price wedge is now a structural seam Anthropic and OpenAI will have to close.
By naming the repo FuckClaude and shipping it in Chinese and Korean, the author telegraphs frustration with Anthropic's pricing wall and packages a subscription-to-API bridge for a non-English developer audience. The 627 stars in a week suggest strong latent demand for treating a Claude Pro seat as a programmable endpoint.
Ships a tool that automates ChatGPT subscription registration and re-exposes it as an API, with Chinese-first docs. The position is that if OpenAI won't offer affordable API access, the subscription flow itself is the API — and 248 stars plus 106 comments show the community treats this as a legitimate infrastructure gap rather than a fringe hack.
Curates a community collection of Anthropic-style Skills bundles, filling the role that a first-party Skills marketplace would have played on launch day. The 489 stars suggest developers want a discoverable registry of prompts-plus-tools packages and are willing to standardize around a third-party index in the absence of an official one.
Argues dzhng/skills 'is what Anthropic's own Skills marketplace should have been on day one,' framing the repo as evidence that Anthropic under-shipped the distribution layer around its own primitive. Groups it with the auth-workaround repos to show a single pattern: the community is filling gaps left by both Anthropic and OpenAI product decisions.
Points to the Chinese and Korean README on FuckClaude, Chinese-first docs on sub2api, and a Western dev-tools founder behind dzhng/skills as evidence of a synchronized global response to the same seam in Anthropic and OpenAI's product line. The claim is that three repos trending in the same week around subscription arbitrage and missing Skills distribution is a signal, not coincidence.
Three repositories trended on GitHub this week, and on the surface they look unrelated. LinXiaoTao/FuckClaude picked up 627 stars with a name that telegraphs its intent. akihitohyh/chatgpt-register-sub2api pulled 248 stars for a tool that automates ChatGPT subscription registration and exposes it as an API. Sandwiched between them is dzhng/skills at 489 stars — a community-curated collection of Anthropic-style Skills, the packaged prompts-plus-tools bundles that Anthropic launched earlier this year.
One name is a middle finger, one is an auth workaround, and one is a legit skills library. But looked at together they're the same story: the AI subscription tier has become the arbitrage layer, and the community is building the pipes Anthropic and OpenAI didn't. FuckClaude and sub2api convert paid Claude and ChatGPT subscriptions into API-shaped endpoints — the same trick Cursor users and cheap-inference wrappers have been running for months, now productized. dzhng/skills fills the other gap: it's what Anthropic's own Skills marketplace should have been on day one, if there had been one.
None of these are one-off jokes. FuckClaude's README is in Chinese and Korean, sub2api's docs are Chinese-first, and dzhng/skills is a Western dev-tools founder's side repo. This is a coordinated pattern across two continents pointing at the same seam in the Anthropic and OpenAI product line.
Start with the subscription-to-API repos, because that's where the honest signal is. Claude Pro is $20/month for effectively unlimited Sonnet and a generous Opus quota. The API bills Opus at $15 per million input tokens and $75 per million output. For any workload heavier than casual chat, the subscription is 10-100x cheaper than the API for the same underlying model. That gap is not a pricing bug; it's a deliberate wedge to drive consumer adoption. But the wedge has grown large enough that the community has built a cottage industry to harvest it.
FuckClaude and sub2api are just the newest entrants. Anthropic quietly ships new detection heuristics — user-agent fingerprints, request-shape anomalies, session-binding checks — and the community ships a patch within days. It's the same dynamic that ran for years around Netflix region unlockers and Spotify family plans, except the arbitraged good is compute, and the mechanic is a Chromium session masquerading as an API client. The Claude CLI's own OAuth token, incidentally, is the exact same lever — anyone who has run the CLI on a server knows the token in `~/.claude/.credentials` can be lifted and reused. That's not a bug the community discovered; it's the front door.
Then there's dzhng/skills. Anthropic launched Skills as a first-class primitive — pre-packaged capability bundles that Claude can invoke: PDF extraction, spreadsheet generation, browser automation. The pitch was that Skills would become a marketplace, discoverable and shareable. Six months in, the primary place developers actually go for a working Skill is a personal GitHub repo. The fact that a community-curated skills library outranks Anthropic's own documentation on trending charts is a quiet indictment of how the official surface has been rolled out — no browse-able registry, no ratings, no install command, no worked examples pinned to a real API doc.
The cross-cutting insight: Anthropic and OpenAI are pushing agentic primitives — Skills, MCP servers, Computer Use, function calling — at a rate the developer surface can't absorb. When the official docs are thin and the pricing model has a 50x arbitrage on it, the community fills both gaps at once. That's what makes these three repos trend in the same week.
Three concrete implications for anyone building on this stuff in production.
First, if you're relying on a subscription-to-API bridge in a real workload, price in the counterparty risk. These tools survive on the exact enforcement gap I described above, and that gap is getting patched every few weeks. When Anthropic ships the check that breaks FuckClaude — and they will — your pipeline goes down at 3am with no support channel. The $20/month savings evaporate the first time a customer-facing feature hard-fails on a Sunday. If your product depends on Claude, use the API, eat the cost, and treat any 10x savings tool as a science-fair project, not infrastructure.
Second, dzhng/skills is worth actually reading. Not because you should trust a random maintainer's skills in production, but because the repo is the closest thing to a working spec for what a Skill *should* look like — folder structure, tool definitions, prompt scaffolding, invocation examples. If you're writing internal Skills for your team, clone it as a template and delete what you don't need. Anthropic's own examples are scattered across three doc pages and a cookbook; this is one place with the pattern you actually copy.
Third, this is a leading indicator for MCP. Model Context Protocol is Anthropic's most important developer bet of the year — a standard for tools to plug into any LLM client. If Skills, which are simpler, are already producing a gray market of community-run libraries and subscription bridges, MCP will attract 10x that activity. Plan your MCP server auth model now, because if you ship one, someone will wrap it in a sub2api-style bridge within a month.
The pattern here isn't going away. As long as consumer AI subscriptions are priced 50x under the API for the same silicon, the arbitrage repos will keep trending, and as long as official developer surfaces lag the feature announcements, the community catalogs will keep filling the vacuum. The interesting question is which of the labs figures out first that shipping a real marketplace — with ratings, revenue share, and a public API to browse it — is worth more than another six months of shipping Skills primitives into a void. Whoever gets there first collects the network effect. Whoever doesn't cedes the discoverability layer to a founder's weekend GitHub repo.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.