Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

Own VPS Didn't Protect Claude Accounts: 51% vs 53% Problem Rates on a Commercial VPN

The advice "rent your own VPS and Claude won't block you" is not supported by a survey of 193 Claude users. Access problems, meaning a block or a KYC (identity verification by documents) request, hit 50.7% of those connecting through their own server and 52.9% of those on a commercial VPN. A stable IP did not separate the groups either: 57% versus 60%. So a VPS should not be treated as a guarantee. Keep a backup tool such as Codex and store project context separately.

AI-processed from Habr AI; edited by Hamidun News
Own VPS Didn't Protect Claude Accounts: 51% vs 53% Problem Rates on a Commercial VPN
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The survey is voluntary, so the findings hold only for the 193 participants
Key metric #2The difference between VPS and VPN is statistically indistinguishable from chance (p≈0.87)
Key metric #3From November 12, 2026, Anthropic takes physical location into account

If you work with Claude or Claude Code from a region where the service is officially unsupported, you have almost certainly heard the "survival recipes": rent your own VPS (a virtual server on which you set up the VPN yourself), don't change your IP, watch for leaks, check the reputation of the address. A Habr author collected 193 questionnaires from Claude users and checked which of this advice the data backs up. The subject is blocks and KYC, meaning identity verification by documents that Anthropic requests for some accounts.

The answer is inconvenient: among those who connected through their own server, access problems occurred for 50.7% (38 of 75), while among users of commercial VPNs they occurred for 52.9% (37 of 70).

The difference is barely visible.

Who Was Asked and How the Numbers Were Counted

The questionnaire had 16 questions: account status, plan, Claude product, KYC history, VPN type, IP stability, exit country, and network diagnostic results. Real IP addresses, emails, card numbers, and documents were not collected. Of the 193 responses, 78 accounts worked without blocks, 54 were restricted, 42 people got a block and gave up on Claude, and another 19 passed identity verification and kept access.

For strict comparisons, the author took two groups with consistent answers. The "problem" group included 84 people: 60 were blocked after KYC, and 24 received a KYC request, did not complete it, and kept restricted access. The "with access" group included 86: 12 passed verification successfully, and 74 were never asked for KYC. Another 23 questionnaires describe ambiguous situations.

An important caveat: the survey is voluntary. A person with a recent ban is more willing to fill in a questionnaire than someone whose access works fine. Therefore the percentages apply only to the participants and do not show the probability of a block for any Claude user.

Own Server and Stable IP

Fisher's exact test (a statistical test of whether differences are random) gave p≈0.87, meaning the difference between a VPS and a commercial VPN is indistinguishable from chance. This does not mean a dedicated server is useless: the questionnaire lumped "own VPS" together with WireGuard and OpenVPN into one option, and the latter two are simply technologies, not a sign of having your own server.

A stable exit IP did not separate the groups either: it was reported by 44 of 77 participants in the problem group (57%) and by 45 of 75 participants with retained access (60%). "Don't know" answers were excluded.

However, one reader described a scenario the questionnaire did not cover. He had an American VPS with an almost year-old unchanged IP, but the VPN would sometimes drop, the internet would disappear, and the computer would come out of sleep. At such moments Claude Code could connect directly, bypassing the VPN. On October 29 he received a KYC request. A connection is not proven, but the practical conclusion is reasonable: a stable server address does not guarantee that the application never went out to the network directly. The application's connection should be blocked when the VPN drops, rather than quietly going through the main channel.

IP Reputation Ratings Give No Guarantees

The author compared two services that assess the network environment: Net.Coffee (Trust Score) and "Plati po vsemu miru" (Pay Worldwide), which looks at the connection from the standpoint of foreign payments. Scores exist for 108 and 83 participants respectively, and for both services for 78.

The Spearman correlation (a measure of agreement: 0 means no relationship, 1 means full agreement) was about 0.46. For 23 of the 78 people, the scores differed by 30 points or more, and for ten by 40 or more.

After excluding ten people with contradictory answers, 68 pairs remained, 18 of which differ by at least 30 points, and the correlation was about 0.47. The author's own connection scored 88 in one service and 40 in the other, while Claude was working.

Medians by group: among working accounts without KYC, 72 points on Net.Coffee and 52 on the second service; among those blocked after KYC, 60 and 40. But among those who passed verification successfully, the medians were only 57 and 42, and the sample was tiny: seven and five ratings. The logic "higher rating means easier KYC" was not confirmed. It is also interesting that for the VPS/WireGuard/OpenVPN category, the median Net.Coffee rating is higher than for commercial VPNs, but this did not affect whether access was retained. The services themselves do not know Anthropic's internal rules, and the measurements may have been taken after the VPN had already been changed.

What to Do in Practice

On November 12, 2026, the updated Supported Regions Policy took effect. Under it, use of Claude by a person physically located in an unsupported region is subject to restrictions regardless of citizenship. This does not mean an automatic ban on first connection, but it can become grounds for restriction. The exit IP's country is only one parameter. The questionnaire did not collect actual country, citizenship, or payment data, so compliance with the rules cannot be reconstructed from it.

The practice for developers and teams: do not build the workflow around a single Claude account. Keep project history and context separate from the service, and have a backup tool ready. The author himself lost his paid Claude Max 20x organization: the subscription was cancelled without a refund, and he switched to Codex.

Story

Key Questions Answered

Who Was Asked and How the Numbers Were Counted?

The questionnaire had 16 questions: account status, plan, Claude product, KYC history, VPN type, IP stability, exit country, and network diagnostic results. Real IP addresses, emails, card numbers, and documents were not collected. Of the 193 responses, 78 accounts worked without blocks, 54 were restricted, 42 people got a block and gave up on Claude, and another 19 passed identity verification and kept access. For strict comparisons, the author took two groups with consistent answers.

What are the key details on Own Server and Stable IP?

Fisher's exact test (a statistical test of whether differences are random) gave p≈0.87, meaning the difference between a VPS and a commercial VPN is indistinguishable from chance. This does not mean a dedicated server is useless: the questionnaire lumped "own VPS" together with WireGuard and OpenVPN into one option, and the latter two are simply technologies, not a sign of having your own server.

What are the key details on IP Reputation Ratings Give No Guarantees?

The author compared two services that assess the network environment: Net.Coffee (Trust Score) and "Plati po vsemu miru" (Pay Worldwide), which looks at the connection from the standpoint of foreign payments. Scores exist for 108 and 83 participants respectively, and for both services for 78. The Spearman correlation (a measure of agreement: 0 means no relationship, 1 means full agreement) was about 0.46. For 23 of the 78 people, the scores differed by 30 points or more, and for ten by 40 or more.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…