Claude Code bans: why they happen, how they are detected, and what actually works
Most explanations stop at "it checks your IP". That is the one thing the documented enforcement did not do. What follows is the structure that the public record actually supports — and the three routes people take when access is gone.
Anthropic's policy lists three conditions in parallel and any one is enough: being physically located in an unsupported region; an entity incorporated or headquartered there; or an entity majority-owned from there. The first is about where you are, not who the account belongs to. A different exit IP does not change the first condition and does not change the other two — which is the real reason VPNs do not help.
Why a VPN does not help
In September 2025 Anthropic updated its restrictions on sales to unsupported regions. The change that matters is scope: the restriction covers entities more than 50% owned, directly or indirectly, by companies headquartered in those regions.
And the first of those three clauses is the one people miss, because it is about a body rather than a company: use while physically located in an unsupported region, including users of otherwise supported entities. A US company and a residence permit do not help while you are in mainland China — a case widely circulated in October 2026 described exactly that combination losing an annual Team plan two days after renewing. Appeals do not reliably help either: Anthropic's own transparency report for the first half of 2026 records about 11.4 million accounts banned and roughly 10.5% of appeals succeeding.
This is why the recurring community complaint — "I was on a US IP and still lost the account" — is not a contradiction. The rule never looked at the IP.
Four layers of signals
| Layer | What it looks at | Can you influence it? |
|---|---|---|
| Account | Registration, payment instrument, corporate ownership | In principle, but not the ownership verdict |
| Request | What the client attaches to each call | No — short of modifying the binary, which is explicitly prohibited |
| Environment | System timezone, preferred language, region format, browser fingerprint | Yes, and this is the only layer worth investment |
| Network | Exit IP country and city, gateway domain | Partly, though the impact is routinely overestimated |
The environment layer is bolded for a reason: it is the only one you control and can verify. It also happens to be where inconsistencies accumulate silently — a machine whose timezone still says Asia/Shanghai while the browser reports a Los Angeles exit needs no sophisticated detection to look wrong.
For the 2026 instrumentation — the timezone read, the domain list, the character substitutions — see How bans are detected.
Find your situation
Three routes that actually exist
If the account is gone or you want out of the dependency, the options are narrower than the internet suggests. They differ in effort, terms exposure and how long they survive.
Run a different harness
A harness is the layer that turns a model into a chat-and-edit agent. A dozen open-source ones are model-agnostic, so no vendor can withdraw access by policy. You give up subagents, hooks and the polish.
Half-life: long Route 2 · Low effortKeep the client, change the brain
Claude Code speaks the Anthropic Messages API and honours ANTHROPIC_BASE_URL. Point that at a compatible gateway and the same CLI drives another model. Ten minutes, one variable to roll back.
Strip the client's built-in controls
Patching or forking the binary. Anthropic's compliance page requires it to run unmodified and forbids disabling built-in authentication. Auto-updates break the result every release.
Half-life: shortest · we do not walk through it| Route | Setup | Keeps workflow | Claude models | Terms exposure | Half-life |
|---|---|---|---|---|---|
| 1 · Different harness | Medium | No | Any | None from Anthropic | Long |
| 2 · Client + other model | Low | Yes | No | Grey | Long |
| 3 · Patch the client | High | Yes | Yes | Explicit | Shortest |
Is it allowed to run Claude Code against another model?
The client stays unmodified, which is what Anthropic's compliance page cares about, and its prohibition targets routing other people's Claude subscription credentials through a service. The backend you route to brings its own terms, rate limits and data handling.
What does ANTHROPIC_BASE_URL actually do?
It replaces the API endpoint the client calls. Auth headers, model names and the tool loop keep working; the backend must accept the Anthropic Messages request shape. An incompatible gateway therefore fails confusingly — the request arrives, the schema does not match.
Which Claude Code alternatives are open source?
Aider, Cline, Roo Code, OpenCode, Crush, Goose, OpenHands, Codex CLI, Gemini CLI and Qwen Code all publish source, mostly under Apache-2.0 or MIT. Confirm each repository's LICENSE before standardising on one, because source-available licences with commercial restrictions also appear in this space.
The four guides
All nine now exist in both languages: three here, and five more below, with the Chinese originals one click away on each page. Every page states what it does not know.
Why Claude Code is not supported in some regions
The service-region list, the 50% ownership clause, and export controls — plus why "I am on a US IP" is not an argument.
In Chinese too 02 · MechanismHow bans are detected
Five layers of signal, the 147-domain list, the 2026 steganographic marker and its timeline — and what it was not.
In Chinese too 03 · RecoveryWhat to do after an account is banned
Tell the situations apart first. Then appeals, refunds, local data rescue, and the fraud patterns.
In Chinese too 04 · RoutesThree routes that actually exist
A different harness, a different model behind the same client, or patching the client — compared by cost and half-life.
On the home pageFive more in English
The rest of the cluster used to be Chinese-only. Each now has an English page at the root path, with the fuller Chinese original linked from its footer.
Banned right after signing up
The four classes of registration signal, graded official versus community-reported, and what to do in the first hour.
06 · MoneyGetting a refund after a ban
Web card, App Store and Google Play are three different routes. A chargeback is not a shortcut.
07 · TravelTravelling without losing the account
Risk depends on direction, not distance: before, during and after checklists.
08 · TracesHow your traces leak
IP and ASN, DNS, WebRTC, browser and TLS fingerprints, behaviour — with commands you can run.
09 · EnvironmentEnvironment consistency checklist
The five layers you can control, item by item, with how to verify each one.
Where CCBaga fits
This site also distributes CCBaga, so the boundary should be explicit: CCBaga does not lift bans and does not modify Claude Code. It inspects and archives local Claude data, switches or resets local identity state, aligns the companion browser's timezone, language and region with the network, and reports whether the measured readings agree.
In the four-layer table above that is exactly one cell: the environment layer. It does not change account ownership, request contents or the exit IP. If the account is already gone, no local tool helps — go to the recovery guide.
Will a VPN prevent a Claude Code ban?
Usually not. The restriction is written around entity ownership, and the 2026 instrumentation read the system timezone and the base-URL domain rather than any network address.
Can a banned Claude account be recovered?
There is an appeal channel, but no published criteria and apparently low success rates. No third party can lift a ban — treat any paid recovery service as fraud.
Does CCBaga unban accounts?
No. It works on local Claude data and the browser profile, never on the account, the network layer or the client binary.