🌉 For mainland China developers · 国内开发者

Claude Code in China.
No VPN. WeChat Pay.

Anthropic's Claude Code CLI is the best agentic coding tool of 2026 — but Anthropic doesn't serve China directly, requires a USD card, and most China-issued cards get declined. AIPower routes Claude Code through Cloudflare global edge, accessible from inside China without VPN, paid via WeChat Pay or Alipay.

10 trial calls · One env var change · Same Claude quality, China-friendly billing

Setup in 3 steps

Claude Code reads ANTHROPIC_BASE_URL and ANTHROPIC_API_KEYfrom environment. Point both at AIPower and you're done.

1

Sign up + top up $5

Get your API key at aipower.me/register. $5 first top-up unlocks Claude Sonnet 4 + Opus 4.6 + adds +100 bonus calls. Pay via WeChat Pay (¥36) or Alipay or card.

2

Set env vars

In your ~/.zshrc or ~/.bashrc:

# AIPower routing for Claude Code
export ANTHROPIC_BASE_URL="https://api.aipower.me/anthropic"
export ANTHROPIC_API_KEY="sk-aipower-..."

AIPower exposes an Anthropic-compatible passthrough endpoint at /anthropic/v1/messages. Claude Code thinks it's talking to Anthropic directly — same request/response shape, same model IDs, streaming SSE supported.

3

Run Claude Code

From any project directory:

# Reload shell to pick up env vars
source ~/.zshrc

# Run Claude Code
claude

# Or pipe a prompt
echo "Refactor this file to use async/await" | claude

Claude Code uses Sonnet 4 by default. Latency from China typically ~50-100ms (Cloudflare global edge), close to native Anthropic from US.

Also works with the official @anthropic-ai/sdk npm package and Python anthropic SDK — set the same base URL + key. Need OpenAI-SDK examples instead? See /cursor · /continue-dev · /cline.

Why this works inside China

🌐 Cloudflare global edge

api.aipower.me hits Cloudflare edge POPs (Beijing / Shanghai / Shenzhen). No VPN. No GFW issues. Same latency as accessing from Hong Kong.

💬 China payment methods

WeChat Pay (¥) and Alipay (¥/USD) supported via Stripe. AI POWER LIMITED is HK entity → no Stripe China-card-decline issues.

📋 商业发票 / Fapiao

Hong Kong commercial receipt on every payment. CN VAT fapiao available for enterprise procurement (contact for details).

🔁 Plus the rest of AIPower

Same API key works for GPT-5, Gemini 2.5 Pro, DeepSeek V3 + R1, Qwen, GLM, Kimi, Doubao, MiniMax. Mix Claude + DeepSeek across your AI tools for cost optimization.

Common questions

Will my Claude Code subscription work if I have one already?

Anthropic's $20/mo Pro plan gives Claude Code a daily quota. AIPower is pay-as-you-go (no monthly fee). If you already have Pro, you can keep using it on USD card directly when Anthropic is reachable, and switch to AIPower only when you're in China or need WeChat billing. Set ANTHROPIC_BASE_URL per-shell or per-project.

Same Claude model? Same speed?

Same model — we route to Anthropic's official API. Speed is comparable (~50-100ms latency from China via Cloudflare edge, vs ~30ms from US direct). Token output speed identical (provider-side).

What about Claude Sonnet's tool use / vision / streaming?

All supported. Our /anthropic/v1/messages endpoint is a transparent passthrough — tool use, multi-turn, vision, streaming SSE all work identically to native Anthropic.

What does it cost?

Per-token billing: Sonnet 4 is $3.60/M input, $18/M output (international pricing). CN-region pricing is higher to fund the cross-border infrastructure — see /pricing for the full table. AIPower's value is China-friendly access plus unified billing across 16 models on one key.

Can I also use Cursor / Continue.dev / Cline this way?

Yes — same approach. Each tool reads either OPENAI_BASE_URL or ANTHROPIC_BASE_URL from env. Set both to AIPower and all your AI tools route through us. Per-tool guides at /cursor /continue-dev /cline.

How is streaming billed accurately?

AIPower forwards SSE chunks unchanged but parses message_start (gives input_tokens) and message_delta (cumulative output_tokens) on the way through. Final token totals are settled when the stream closes. This means: (a) you see Claude's output token-by-token with no buffering delay, and (b) billing matches what Anthropic counted. Code: github.com/yangyongfee-tech/aipower-examples or read packages/api/src/routes/anthropic.ts.

Is this production-safe?

Yes for individual / small-team usage. For mission-critical production, run on AIPower's Team or BYOK plan — both add per-key budgets, encrypted upstream key vault, and SLA support. Worker-tier shared rate limits apply to free + pay-as-you-go accounts (well above any single-developer use).

What if Anthropic rate-limits AIPower?

Today: api.anthropic.com is our single upstream and we operate at a usage tier that's well above current load. If we ever approach limits, we'd add AWS Bedrock or GCP Vertex Anthropic endpoints (same Claude models on different infra) — they're built into the architecture but not yet active. Live status at /status with daily per-provider uptime.

What if I want to switch back to direct Anthropic later?

Trivial — just unset ANTHROPIC_BASE_URL. The Anthropic SDK falls back to api.anthropic.com automatically when no base URL is set. No code changes, no migration. Your existing Anthropic API key takes over immediately.

Does this work with Claude Code's tool-use loop (file edits, bash, search)?

Yes. Claude Code makes multiple round-trips per agentic task (read file, propose edit, run bash, etc). Each round-trip goes through our passthrough independently — same auth, same accumulated billing. The forwarder is transparent to tool-use payload shape. If you hit any tool-use edge case, email frank@aipower.me with the failing request and we'll debug — early adopter feedback shapes hardening priorities.

Start using Claude Code in China — $5

First top-up adds +100 bonus calls. Pay via WeChat in 30 seconds. No VPN setup.

Deep dive