vm2api: after burning 30 Claude accounts in testing, a relay gateway picked up 51 stars in 1,780 days

On Sep 16, 2026, a GitHub account less than half a year old published a repo and posted on linux.do, the Chinese developer forum, under the title "We burned through 10+ max 20 pro accounts in testing." The repo turns Claude and ChatGPT monthly subscriptions into API access. The space already had two incumbents with 40-50K stars each. 51 days later it had about 1,780 stars and 451 forks, a fork-to-star ratio of 25%. In a crowded niche, it carved out its spot with one line: "anti-ban." Starting Nov 12, that line collided head-on with Anthropic's new rules.

A headline paid for with accounts

The author's linux.do ID is mci77777, and on GitHub it's dofastted. The account was registered in April, and its profile links only to a Telegram group. The full launch post title was "Open-source collaboration. Virtual machine kernel claude codex to api. New architecture. Anti-ban. We burned 10+ max 20 pro accounts in total." The body said: "My claude pro accounts have lasted 1 months, and max lasted 2 weeks," and "the usage is no different from cpa sub2api."

A Max subscription costs 100 to 200 USD a month, and Pro is 20 USD. Burning 30 accounts means the author paid roughly a couple thousand USD (estimate) to test on behalf of readers. For people on L-site who get banned all the time, that's more convincing than any feature list.

Five days later, the launch post was gone. The author opened a new "vm2api discussion thread" to explain: "The previous post is dead. Open-source promotion can't include closed-source content, so I'm not posting the link now… After version 1.3, the core code will be closed." That thread got 1.7K views and 39 likes, and people kept heading to GitHub anyway.

What it is

A gateway you host yourself. You put in your Claude Pro/Max account (ChatGPT/Codex is supported too), and it gives you a standard OpenAI or Anthropic format API that multiple people can share, with issued keys and billing.

On this point it's the same as the veterans. The difference is the approach: each account gets its own virtual machine or Docker container (called a "slot"). Inside, an unmodified official Claude Code process runs continuously. Hardware fingerprints (motherboard info, MAC address, machine-id) are made to look like a real machine. Each slot uses one exit IP, and it won't start without an IP bound. Before a request enters a slot, it also passes a "risk audit" that blocks anything likely to trigger a ban locally. When usage nears the official 5-hour or 7-day limits, it trips a circuit breaker automatically.

A one-click script plus Docker Compose handles installation. The README homepage reads: "Fully isolated VM/container environment + official client process running continuously + realistic hardware fingerprints + one dedicated network exit per slot + pre-entry anti-ban interception."

Why this one

The space is already crowded. The "subscription-to-API" niche is called "2api" in Chinese circles. The leaders are CLIProxyAPI (called CPA in the community, launched 2025-07, 57,692 stars) and sub2api (launched 2025-12, 45,522 stars), each 25 to 30 times bigger than vm2api. Below them are claude-relay-service (12k stars), AIClient2API (9k stars), and a pile of small projects that only support Codex.

The veterans have a structural weakness. CPA and sub2api both imitate official clients at the HTTP layer: they rewrite request headers and simulate OAuth. In Oct 8 and Oct 24, two waves of mass bans were discussed on L-site, and someone posted that a new Claude Code version added markers specifically to detect third-party gateways. At that point, "forged request headers" were seen as "exposed in a minute."

vm2api was built one layer differently: no forgery, it just runs the real client. One user summed it up in a post: "Unlike sub2api, which fakes header info, this uses the real Claude Code… more like a foreigner who's physically there." The idea fits in one sentence, and it launched right when ban anxiety was at its worst.

The author stayed active. 51 days, 2,044 commits, 320 versions. Between Nov 5 and 6, they shipped 3 versions in one day. The thread got replies almost every day, and about 20 outside developers submitted PRs. Users felt someone was there when things went wrong.

Switching costs almost nothing. The external interface matches the old tools, so existing users can switch by changing the address.

Where users come from

Nearly all of it happened on linux.do, in one thread:

The reverse-proxy log thread was the best ad: one user 10-22 wrote there, "After getting banned 20 times… using vm2api finally lasted a whole day." Even the negative threads were pulling people in.

Off-platform coverage was sparse: one GitCode blog post (10-16,1,627 reads), a YouTube Shorts clip that looked auto-reposted, and a few trend aggregator sites. X had no discussion in the last 60 days, and no media covered it.

Where the money comes from

Open source doesn't mean free to commercialize. The license is custom: source is public, commercial use is prohibited. For commercial use you contact @VM2API on Telegram to buy a license. The changelog also mentions a paid core called CRAG, which issues registration codes per VM, in 1-month, six-month, and 1-year tiers. The author also runs a hosted gateway that sells keys.

The cleverest part is the sponsor slot at the top of the README: ProxyLane static residential IPs, "dedicated IP per slot, from $3.90," with discount code VM2API30 for 7% off. "Each slot must bind one IP" was originally an anti-ban design, and it doubled as a reason to sell IPs. Revenue hasn't been disclosed.

What it doesn't do well, and what it can't stop

• Results can't be verified. Anti-ban claims rest entirely on users reporting how long their accounts lasted, both good and bad. During the 10-24 ban wave, the author admitted losing 2 Max accounts. Some say cache hit rates are poor and Pro quota drains faster than with CPA, and others urge people "never update."
• Heavy. One VM per account plus a residential IP per slot means deployment and running costs are noticeably higher than a single-process setup like CPA.
• The rules changed. Anthropic's new usage policy, announced 11-02 and effective 11-12, explicitly bans two things: packaging Pro/Max subscription quota as paid API access, and letting third-party tools impersonate official clients (per ChainCatcher 11-02). Those are exactly vm2api's selling points.
• Cooling. Over the last 7 days, +144; in the first 7 days, +204. In November it's gaining only 6–30 a day.

Others have already copied it: mrhong93/vm2api maintains its own 2.x distribution; someone on L-site says they forked it and added Grok; someone on macos.do built a sub6api with the same idea, and the author admitted "someone already did it." The more interesting signal is the periphery: DUNHKpcc/Claude-Guardian was created recently, reached 10-15 stars in three weeks, and is still gaining 1,761–487 per week. It shows "anti-ban" itself is a hot need, and whoever explains it clearly gets the stars.

If it were you

C. Don't build another vm2api. This niche is working against platform rules: the top CPA and sub2api have 40–50K stars, the rules explicitly banned this model as of 11-12, late followers get a short window, and the project could be banned or pulled at any time, which also looks bad on a résumé. What's worth learning is how it positioned itself in a crowded field.

Good fit for: developers who want stars in a niche that already has big projects, who can afford real money and real time for testing, and who are willing to stay in one community long term and reply daily.

Bad fit for: anyone building "subscription-to-API" or "anti-ban" tools (crowded niche, tightening rules, and it could go to zero overnight); anyone who cares about their open-source record or wants to use the project for job or partnership leads; anyone without a regular community where their posts get seen.

How to do it: apply the playbook to a compliant niche. ① Find a structural weakness in a top open-source project that "a small change can fix," and state it in one sentence (a pattern like "it fakes request headers, I run the real client"); ② use a costly real-money test as your headline (how many accounts burned, how many hours run, how much money spent); ③ open a continuously updated "test log" thread where users post results; ④ ship releases often in the first two months and reply to threads yourself.

How to get users: launch in one vertical Chinese community with real pain points and dense discussion (like linux.do), not Hacker News or X. If the launch post gets deleted, continue in a "discussion thread." Drive traffic by replying in related question threads. Reply to negative threads too, since that's exposure.

Deadline: for the vm2api route specifically, the effective date of the 11-12 rules is the deadline, and how long it survives depends on enforcement. For the "change one layer" playbook there's no fixed window, but you have to move before the top projects patch that weakness themselves.

Signs you misread it: ① In the month after 11-12, ban discussion for vm2api and similar projects doesn't noticeably rise, and stars keep recovering. That would mean platform enforcement is loose and the route lasts longer than expected. ② The reverse: if CPA or sub2api adopts a "real client + isolation" model themselves, and vm2api's star growth stalls below 2,000, the "change one layer" position can be erased by one release from the leader, and the playbook is worth less.

Data and sources

• GitHub: dofastted/vm2api's README, LICENSE, and changelog; stars, forks, commits, and version count (2026-11-06); daily star counts (logged daily from 10-17).
• linux.do: the launch post (9-16, archived at ifeed.cc), the discussion thread (linux.do/t/topic/2931202), the reverse-proxy log thread (linux.do/t/topic/2959633), the reply thread (linux.do/t/topic/2980881), and the ban wave thread (linux.do/t/topic/3068174).
• macos.do: sub6api discussion thread (macos.do/t/topic/266).
• Anthropic policy: ChainCatcher 2026-11-02 "Anthropic updates Claude policy, bans reselling subscriptions and disguising clients."
• Competitor star counts: CLIProxyAPI, sub2api, claude-relay-service, AIClient2API, Claude-Guardian (GitHub, 2026-11-06).
• Unverified: the direct cause of the single-day peak on 10-14; the exact reason the launch post was deleted; CRAG's price and sales; the makeup of the star accounts.