Three ways to reach Claude from mainland China: official, relay, mirror

Tutorials Published 2026-09-19 Updated 2026-09-26 92 views API Compass
There are really three paths to using Claude with Claude Code from mainland China: a direct official connection, a third-party relay, and a web mirror. The differences that matter are not the price.

What the three options actually are

OptionWhat it isWhat you need
Direct officialCalling the Anthropic API yourselfOverseas network access + an overseas payment method
Third-party relayBuying credit from a relay service and going through their resale channelCNY payment, and a base_url to configure
Web mirrorSomeone wrapping the claude.ai web client as a siteJust a browser — but no API

Note the last row: a web mirror gives you no API key. It just lets you use the web version. If what you want is Claude Code, Cline or Roo Code, a mirror does not help — all of those need an API endpoint.

So the real choice is between the first two.

Option one: direct official

Process: register an Anthropic account → attach an overseas card → top up → get a key → set environment variables.

The advantages are full capability and no intermediary: every new model and feature is available immediately. The drawbacks are equally clear:

  • You need a stable overseas network. A network wobble shows up as Claude Code hanging.
  • You need an overseas payment method.
  • Account risk controls are strict; unusual usage can get you limited.

If you have stable network and payment access, this is the least friction.

Option two: third-party relay

Process: register and top up at the relay service → take their key and base_url → configure your client.

The client configuration comes down to a few environment variables:

ANTHROPIC_BASE_URL=https://the-address-the-service-gives-you
ANTHROPIC_AUTH_TOKEN=the-key-the-service-gives-you

The advantages are that it works on an ordinary local connection, you pay in CNY, and the barrier is low. The risks concentrate in three places:

  1. Channel stability. As noted above, low prices often come from resold subscription quota, so a vendor ban affects you too.
  2. Whether the capability is complete. Some relay services proxy only the chat endpoint and do not forward tool use. The result is Claude Code that can chat but cannot edit files — by far the most common "it works but not really".
  3. Whether the model is downgraded. Especially at services priced noticeably below their peers.

How to check: after connecting, give Claude Code a real task that requires reading and writing several files. If it only prints text and never calls a tool, the endpoint does not support tool use. Switch.

Option three: web mirror

It suits exactly one scenario: you only want to chat with Claude, not build anything.

The moment you need API calls, batch processing, or integration into your own program, a mirror is useless.

Mirrors are also usually built by reverse-engineering the web client, so they are the least stable, and shared accounts raise a privacy problem — your conversations may sit in the same account as other users'.

How to choose

  • You just want Claude Code running for personal projects → find a relay service that explicitly supports Claude Code, top up a small amount first and confirm tool use works. Services in our Claude category state their compatibility.
  • Team production use, stability required → prefer a direct official connection, or a provider that can issue invoices and has an identifiable operating entity.
  • Chat only → a mirror is enough, but do not put anything sensitive through it.

One cost item people miss

Most people compare rates and ignore the hidden cost of retries.

Relay services throw the occasional 500 or timeout, and the client retries automatically. Those tokens are billed too. A service at rate 0.3 with a 10% failure rate can end up costing more than a stable service at rate 0.4.

That is why we insist on showing uptime and latency. A cheap rate is only actually cheap if the service is stable.

Was this useful? If you have run into something the article does not cover, tell us and we will verify it and update the text.

Related articles

Video generation behind a relay: three API shapes, four billing traps, one acceptance checklist

2026-10-19

MCP through a relay: what works, what never will, and how to test it in ten minutes

2026-10-19

A relay selection decision tree: working back from your scenario to a site

2026-10-07

Logging and reconciliation: attributing token spend to the project level

2026-10-06

Migrating suppliers gradually: replacing an API site without interrupting service

2026-10-06

Build your own probe: twenty lines of script to monitor your API suppliers

2026-10-06

The models this article mentions are available on these services

Filtered by the model families the article mentions and ordered by overall score and uptime. This is not a recommendation. Click through to see each service's uptime, latency and endpoint verification.

See the full list in the directory, or browse by model family.