By building and naming the repo 'FuckClaude,' LinXiaoTao takes an explicitly confrontational stance: Anthropic's per-token API pricing is untenable when a flat-rate subscription exists. The proxy converts a Claude Pro/Max web session into an OpenAI-compatible endpoint, arguing developers shouldn't pay twice for the same underlying model.
The chatgpt-register-sub2api project applies the same arbitrage logic to OpenAI: wrap a ChatGPT subscription as an API to bypass metered billing. The 204 stars and 94 comments suggest broad developer agreement that subscription-to-API conversion is a legitimate workaround, not just a Claude-specific grievance.
Argues the gap between $20/month subscriptions and API costs that can burn $20 before lunch has grown so wide that entire GitHub projects now exist just to bridge it. Frames FuckClaude's star count as a 'vote' rather than a bookmark — developers signaling that the pricing model itself is the problem.
The dzhng/skills catalog piggybacks on Anthropic's official Agent Skills feature, which runs inside the Claude app under the subscription tier. Rather than bypassing Anthropic's system, it maximizes what subscribers can do within it — treating Skills as the sanctioned path to agentic behavior without hitting the API meter.
The science-skills repo extends the same pattern to a specialized domain, showing the Skills-as-community-library approach is generalizing. 197 stars and 51 comments indicate developers see curated Skills libraries as a durable way to compound subscription value without violating terms of service.
By releasing LongCat-2.0 as an open-weight LLM from a Chinese food-delivery giant, Meituan sidesteps the pricing debate entirely — developers can download and run it locally with zero per-token cost and no subscription. The 248 stars signal appetite for a third path where the frontier-lab pricing structure simply doesn't apply.
Reads the three trending repos as a single signal: subscription arbitrage, Skills exploitation, and open-weight defection are all responses to the same underlying tension. Argues Anthropic's revenue model becomes 'an all-you-can-eat buffet for anyone willing to run a proxy,' implying the current pricing structure is unstable and will force either enforcement, tier restructuring, or margin compression.
Three unrelated repos hit GitHub trending this week and, read together, they tell a single story. LinXiaoTao/FuckClaude cleared 487 stars on the week — a proxy that dresses a Claude Pro/Max web session up as an OpenAI-compatible HTTP endpoint. dzhng/skills pulled 251 stars — a community Skills catalog that piggybacks on Anthropic's Agent Skills feature so subscription users get agentic behavior without hitting the API meter. And meituan-longcat/LongCat-2.0 landed at 248 stars — an open-weight LLM from Meituan (yes, the Chinese food-delivery giant) that you can just download and run.
Three different repos, three different countries of origin, one shared premise: don't pay per token if you don't have to.
The naming alone is a signal. "FuckClaude" is not a repo people star to bookmark — it's a repo people star to *vote*. The README makes the pitch explicit: convert your existing Claude subscription into an API surface, then point Cursor, Cline, or your own agent framework at it. Rate limits are the subscription's rate limits. Pricing is the subscription's pricing. Anthropic's revenue model becomes an all-you-can-eat buffet for anyone willing to run a proxy.
The economics here are not subtle. A Claude Pro seat is $20/month. A Max seat is $100–200. An engineer running a coding agent through the Anthropic API on a normal workday will burn through $20 of tokens before lunch — Sonnet 4.5 at ~$3/M input and $15/M output, with agents that re-read the same context every turn, is not a cheap habit. The gap between subscription pricing and API pricing is now wide enough that entire GitHub projects exist just to bridge it.
dzhng/skills is the more polite version of the same move. Anthropic shipped Agent Skills as a way to teach Claude persistent, reusable capabilities — think tool wrappers with markdown scaffolding. Skills work in the Claude app, which means: subscription. A community library of skills makes the subscription meaningfully more powerful for exactly the workloads people would otherwise reach for the API to run. Every skill someone contributes is one more reason not to pay per token.
LongCat 2.0 attacks the same problem from the opposite side. Meituan is not a foundation-model company — they deliver dumplings. But they have GPUs, they have engineers, and open-weight releases from Chinese labs (DeepSeek, Qwen, Kimi, MiniMax, and now LongCat) have become monthly events. The 2.0 release includes multiple sizes with permissive-enough licensing that a small team can self-host on rented H100s and pay wholesale compute prices instead of retail token prices. When the third-largest food delivery app in China ships a competent LLM in the same week that a proxy called FuckClaude trends, the market is telling you something.
The community reaction is worth reading. HN and Reddit threads under all three repos share a specific tone — not evangelism, not benchmark obsession, just tired math. "I was paying $400/mo in Anthropic bills." "Cursor's usage-based pricing burned $80 in a weekend." "Switched to Max + this proxy and my bill is fixed." The people starring these projects are not researchers. They're working engineers who watched their AI bill outrun their AWS bill.
Anthropic and OpenAI are aware. Both quietly tightened subscription terms of service in 2025 to prohibit programmatic use — the FuckClaude README even links to Anthropic's TOS, essentially daring them to enforce it. The enforcement problem is that these proxies look like a browser. Serving a takedown to every fork is whack-a-mole. Serving a ban to every user is churn.
If you're a working developer running agents, the calculus has shifted. Three months ago the default answer was "pay for API access, it's the professional choice." Today the default is "which subscription, and which proxy." That's not a moral judgment — it's what the trending page is telling you the herd is doing.
A few practical reads:
If your team is already spending real money on Claude API, at least benchmark a Max seat + proxy against your current bill. Not to switch — to know the number. If the delta is 10x, your finance team is going to find out eventually and you'd rather be the one who ran the numbers first.
If you're building on top of Anthropic or OpenAI as a platform, understand that a nontrivial slice of your "API users" are running proxies against subscription accounts. Your usage graphs are lying to you about the true price sensitivity of your customer base. The customers most vocal about token costs are the ones already gone.
If you're evaluating open-weight models, LongCat 2.0 is worth actually testing rather than dismissing on brand. The gap between frontier closed models and the best open weights has been narrowing on coding and tool-use benchmarks for six months straight. "Meituan" sounds unserious until you look at the eval numbers. Kimi K2 sounded unserious too.
And if you maintain a project that costs money to run per user — you now have a template. Skills libraries, proxy layers, and self-hostable model wrappers are the three shapes of "escape the meter." Expect all three patterns to keep showing up.
The end state of this is either a pricing correction — closed labs offer real per-seat plans that acknowledge power users exist — or an enforcement war that turns every proxy repo into a Streisand event. The former is more likely, because the labs need developer mindshare more than they need the marginal revenue from throttling their most engaged users, and because the open-weight ecosystem is now good enough that pushing developers away has a real destination. Watch what Anthropic does with subscription tiers over the next quarter. If a Claude Team plan with a real per-seat token allowance appears, that's the tell. If instead the TOS gets more aggressive and proxy repos start getting DMCA'd, the arbitrage era is not over — it's just moving to smaller, less discoverable Git hosts.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.