Short answer: Cursor works from mainland China in 2026, but not smoothly — subscribing usually takes a mainland-issued AMEX card, a family member or colleague abroad on a Teams plan, or a reputable virtual card, and latency to Cursor's backend varies day to day. For heavy agent work, Claude Code pointed at Router One's China-reachable endpoint sidesteps both the card and the latency problem; many developers run both.
Cursor is great. Cursor from China is a different story. The editor itself runs locally, but every meaningful interaction depends on Cursor's backend, and that backend assumes a working foreign card and a stable, low-latency connection to a US data center. Both assumptions break the moment you try to subscribe from Beijing.
This guide is honest about what works in 2026 and what does not. We cover the three subscription paths Chinese developers actually use, what Cursor's network problems look like, and the case for switching to Claude Code through Router One when Cursor's friction outweighs its benefits.
What Breaks From China
Three concrete failures show up:
Subscription. As of September 2026, Cursor sells Pro, Pro+ and Ultra plans for individuals and Teams for companies; each plan includes a set amount of model usage, and on-demand usage beyond that is billed in arrears (current terms on cursor.com/pricing). All of it is billed through Cursor's own card processor, with the usual fraud rules. Cards issued in Mainland China — Visa, Mastercard, UnionPay — are rejected at a rate high enough that "use your existing card" is not a real path for most people.
Network. Cursor's backend (api2.cursor.sh and friends) lives behind Cloudflare US edges. Mainland ISPs reach it sometimes, and latency swings widely by ISP and time of day. The autocomplete model in particular is sensitive to latency — when the round trip stretches, suggestions arrive after you have already typed the next character, and you lose the type-and-accept rhythm the product is built on.
Routing lock-in. Unlike Claude Code or Codex CLI, Cursor cannot be pointed wholesale at a custom endpoint. The "OpenAI API key" setting lets a custom key bill the chat and Agent-panel models you pick through Router One, but Tab autocomplete, Auto and Cursor's own Composer models stay on Cursor's built-in models, and file edits in Agent mode are worth trying, not guaranteed. Cursor's help page (checked 2026-09-29) adds that requests made with your own key are still routed through Cursor's servers, so a custom key does not remove the network problem described below. Most users want the full product, not the BYO mode. If the BYO mode is what you actually want, a VS Code agent extension such as Kilo Code takes the Router One key directly — the setup guide has the fields.
We've covered the broader Cursor vs Claude Code tradeoff in Cursor vs Claude Code: which AI coding tool for which workflow.
Subscription Path 1: Mainland-Issued AMEX
Several Chinese banks issue AMEX cards with USD support (for example CMB, ICBC and Bank of China). Reports suggest Cursor's checkout accepts AMEX more readily than UnionPay, but outcomes vary by bank, card tier and month, and there are no reliable public success rates — treat it as worth a try, not a sure path.
If your card is rejected on first attempt, do not retry the same card immediately — that increases the fraud score. Instead, try a different card if you have one, or move to path 2 or 3.
Subscription Path 2: Family or Colleague Abroad
The most stable path is to subscribe through someone with a US/Singapore/Canada/UK card and use a Cursor Teams plan to share access. Teams is explicitly multi-seat, and a small team paying through one billing entity is within ToS.
This works well for two patterns:
- A relative or friend in the destination country who fronts the subscription and you reimburse domestically.
- A small dev shop with a billing entity registered abroad and team members in China.
The downside is admin overhead — invoices need to flow somewhere, the billing email is on the foreign side, and adding/removing seats requires the foreign account holder.
Subscription Path 3: Reputable Virtual Cards
Several services issue prepaid Visa/Mastercard usable for Cursor's card billing. Examples: WildCard, Nobepay, OneKey Card. Acceptance rates against Cursor specifically fluctuate as its processor updates fraud rules. "Works most months" is the honest summary.
Two important caveats:
- Pick a service with multi-month track record. Newer services have higher decline rates because the processor has not yet seen patterns from those BINs.
- Do not pick the cheapest. Cards in the $5-15 issuance range often hit decline early; cards in the $20-30 range tend to survive longer.
This path is the most fragile of the three.
Once You're Subscribed: The Network Problem
A working subscription does not solve latency. Cursor's edge is in US-East and US-West regions; the public ingress is fronted by Cloudflare's global anycast, which means Mainland connections route through whichever Cloudflare PoP your ISP decides — typically Hong Kong, Singapore, or sometimes Tokyo, with occasional drift to LA or Seattle on bad days.
In practice the round trip depends on your ISP, the PoP your traffic lands on and the time of day: a good route to a nearby PoP feels close to local, while peak hours on a congested route bring long round trips and timeouts. Router One does not measure Cursor's endpoints, so we publish no Cursor latency figures.
Autocomplete needs sub-200ms to feel responsive; chat tolerates 500 ms; agent mode tolerates whatever it has to. So Cursor is usable from China most of the time, but the autocomplete experience that the product is famous for is degraded compared to a developer in San Francisco.
There is no clean fix for this from the user side because Cursor controls the routing. The options are: live with it, run a low-latency proxy in HK/SG that fronts Cursor (complicated and against Cursor's network ToS in spirit), or use a different tool when latency matters.
When Claude Code Through Router One Beats Cursor
For tasks where Cursor's autocomplete latency is the bottleneck, Claude Code is a better fit because:
- It runs in the terminal, so latency to the agent matters but not to your keystrokes — you don't feel it the way you feel autocomplete latency.
- You can point it at an Anthropic-compatible endpoint such as Router One, which measured 110-130 ms p50 from Beijing, Shanghai and Shenzhen in the published China latency benchmark.
- Billing is RMB through Alipay — no card subscription friction.
The configuration is two environment variables — full setup in Claude Code setup guide and the broader Claude Code in China guide.
The honest tradeoff: Claude Code is not Cursor. The interaction model is different — you write prompts and the agent writes files, rather than tab-tab-tabbing through autocomplete. Many developers in China end up using both: Claude Code through Router One for autonomous tasks (refactors, multi-file features), Cursor for inline edits when the latency is good enough that day.
A Hybrid Workflow That Works
The most common pattern among China-based teams in 2026:
- Cursor for editor-native interaction — Cmd-K, inline chat, file navigation. Use Cursor's cheapest paid tier (Cursor Pro, $20/mo) and ignore the agent feature when network is poor.
- Claude Code in a terminal tab for autonomous work — refactors, large features, debugging sessions where the agent runs tests and iterates.
- Router One as the LLM backend for both — Claude Code points at it directly; Cursor uses it via the BYO OpenAI key option for fallback when Cursor's hosted models are slow. The Cursor integration guide and the Cursor China setup page walk through the BYO override step by step.
This split lets you keep Cursor's editor ergonomics where they shine and bypass them where they don't. See Router One vs OpenRouter for China for why Router One specifically suits this hybrid setup better than other gateways.
Cost Snapshot
| Setup | Monthly cost (heavy use) | Notes |
|---|---|---|
| Cursor plan only | Plan fee, plus on-demand usage once the included amount runs out | Autocomplete still slows on bad network days from China |
| Claude Code through Router One only | $30-150 (pay-per-token) | Wallet billing or subscription credits, no card friction |
| Hybrid: Cursor Pro + Claude Code via Router One | Cursor plan fee + Router One usage | Best ergonomics, both billing paths needed |
For light users (under 1 hour/day) Cursor Pro alone wins on price. For heavy users (full-time AI-assisted) the hybrid wins on overall productivity, despite the higher cost.
FAQ
Can I use Cursor's "BYO OpenAI key" with Router One's key?
Yes. In Cursor settings, set the OpenAI API key to your Router One key and the base URL to https://api.router.one/v1. The chat and Agent-panel models you pick then run through Router One, while Tab autocomplete, Auto and Cursor's own Composer models stay on Cursor's built-in models. OpenAI has announced a proposed November 12, 2026 end to its contract supplying OpenAI models to Cursor; the Cursor integration guide covers what that means for this setting.
Will my Cursor account get banned for using a virtual card? We have not seen widespread bans tied specifically to virtual card payment, but the fraud signal can lead to ad-hoc account flags. Using AMEX or family-account paths is lower risk.
Why doesn't Cursor let a custom key drive Tab, Auto and Composer? Cursor's value is partly its routing decisions and proprietary autocomplete. Opening those up to arbitrary endpoints would commoditize the product. This is a known constraint of using Cursor specifically; other tools in the space (Cline, Kilo Code, Claude Code) make different tradeoffs.
Is there a GitHub Copilot equivalent that works better in China? GitHub Copilot has the same network and payment issues as Cursor, with fewer mitigations. One mitigation has since appeared: Copilot's Custom Endpoint provider lets Copilot Chat and agent sessions run on your own key — the GitHub Copilot setup guide has the fields — though inline completions stay on GitHub's service. Continue.dev used to be the closest "Copilot-style autocomplete with full BYO endpoint" setup, but it is no longer maintained: its README (checked 2026-09-29) marks the repository read-only after a final 2.x release on 2026-06-19. Existing installs still take a Router One base URL; for a maintained extension that takes the same base URL, key and model ID, look at Cline or Kilo Code, and check each tool's docs for which features, such as inline completion, run on a custom provider.
What about Cursor's voice mode and image features? Voice and image inputs are part of Cursor's hosted features and depend on Cursor's backend. They work when the network is good and degrade with the network when it isn't. There is no proxy fix for this from the user side.
Should I worry about my code being uploaded to Cursor? Cursor has a Privacy Mode setting that prevents your code from being used for training. Codebase indexing still happens server-side regardless. Claude Code can run locally while sending model requests to Router One and an upstream provider; this is not a fully local data path. Router One does not retain prompt or model-response bodies from direct API calls. It records request metadata for billing, usage tracking, and troubleshooting. Playground saves conversation history so users can revisit and continue their chats.
Conclusion
Cursor from China in 2026 is workable, not seamless. Mainland AMEX or a family account get you subscribed; the network varies day-to-day. For the heaviest workloads — refactors, agent loops, anything where autocomplete latency would slow you down — Claude Code through Router One is genuinely faster and removes the card friction entirely.
The pragmatic answer for most developers is to use both. For configuration help see the Claude Code setup guide; for the underlying network and routing story see AI model routing explained.