判断中转站，我看四点：

第一，模型目录是不是按分组算的。 很多站宣传「几十个模型」，但那是全站聚合数字，单个 Key 能调的只是它所在分组的目录。建错分组，用 Claude 时直接报 model_not_found，新手最容易栽在这。

第二，价格是否公开明码。 敢把每个模型每百万 token 的输入/输出价挂在公开页、不用登录就能看的，至少计费口径可核对；只写「低至几折」的要小心。

第三，有没有基本的防滥用。 注册要邮箱验证、有一定的风控，说明站方在控成本，这种站反而活得久；什么门槛都没有、撒钱无上限的，多半跑得快。

第四，支付和起充门槛。 起充越低试错成本越低，先小额验通自己的工作流，再谈充多少。

拿 ZeroShell 当对照样本：接口是 OpenAI 兼容的 https://zeroshell.top/v1，同时有 Anthropic 原生 /v1/messages，所以 Claude Code 用官方分组的 Key 配 ANTHROPIC_BASE_URL=https://zeroshell.top 就能跑（这步实测通过）。分组情况如实说：低价组 27 个模型（DeepSeek、GLM、GPT、Grok、Kimi、Qwen，没有 Claude），官方组 22 个含 Claude 全系，46 是全站聚合数。价格举例：gpt-5.6-luna 输入 ¥0.075/百万，gpt-5.6-sol ¥0.375/¥2.25，claude-sonnet-4-6 ¥2.025/¥10.125，gemini-3.1-pro-preview ¥0.75/¥4.5，全在公开价格页。注册约 1 分钟、要邮箱验证，新号送 ¥1，邀请双方各 ¥1。

短板也明说：目前只能支付宝充值（$1 起，¥1=$1 额度），站子新、规模小，我不承诺什么可用性指标。建议就用送的 ¥1 把你自己的流程验一遍，通了再决定。问题反馈：support@zeroshell.top。

The two endpoints you need to know

A relay typically exposes:

• POST /v1/chat/completions — the OpenAI-compatible endpoint. Base URL form: https://zeroshell.top/v1.
• POST /v1/messages — the Anthropic-native endpoint. This is the one Claude Code actually requires.

That second endpoint is the real compatibility test. Plenty of relays advertise "Claude support" but only translate OpenAI-format requests; Claude Code sends Anthropic-format requests and fails against those. Before paying anything, verify the relay documents and serves /v1/messages. On ZeroShell both formats are served, and both have been exercised with live requests — Claude models respond over the OpenAI format and over /v1/messages.

Claude Code configuration

Claude Code reads its endpoint from environment variables:

export ANTHROPIC_BASE_URL=https://zeroshell.top
export ANTHROPIC_API_KEY=<your key>
claude

Note the base URL has no /v1 suffix here — Claude Code appends /v1/messages itself. Model names come from the relay's own catalog, for example claude-sonnet-4-5, so copy them from the relay's model list rather than typing from memory; a wrong suffix is the most common self-inflicted error.

The pitfall: one account, several catalogs

Here is the part most relay documentation glosses over. On ZeroShell, an API key belongs to a group, and each group has its own model catalog:

• A low-cost group key sees 27 models: DeepSeek, GLM, GPT, Grok, Kimi, and Qwen families. No Claude models at all.
• An official group key sees 22 models, including the full Claude lineup.

The site advertises 46 models — that figure is the aggregate across all groups, not what any single key can call. If you create a low-cost key and point Claude Code at it, every request returns model_not_found, and the error message blames the model rather than your key's group. The fix takes thirty seconds (create a key in the correct group), but only once you know groups exist. One account can hold keys in several groups, and the balance is shared, so the practical setup is: an official-group key for Claude Code, a low-cost key for bulk OpenAI-format work.

When evaluating any relay, ask this question first: is the advertised model count per-key or per-account? The answer determines whether your first integration attempt works.

Other OpenAI-compatible tools

For tools like Continue, Aider (in OpenAI mode), or chat clients such as Cherry Studio and NextChat, configuration is the standard trio:

• Base URL: https://zeroshell.top/v1
• API key: a key from the group that contains your model
• Model: the exact catalog name, copied, not recalled

A 30-second smoke test before wiring up any tool:

curl https://zeroshell.top/v1/chat/completions \
  -H "Authorization: Bearer <your key>" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-5-mini","messages":[{"role":"user","content":"ping"}]}'

If this returns a completion, every OpenAI-compatible client will work. If it returns a model error, suspect the group/catalog mismatch above before suspecting the endpoint.

What it costs

Relay pricing is per million tokens, published on the pricing page. Current listed prices:

Worked example: a heavy personal month of 20M input and 5M output tokens costs $19.25 on gpt-5-mini, $85 on gemini-2.5-pro, and $135 on claude-sonnet-4-5. That spread is exactly why the group system exists — route summarization and boilerplate to the cheap catalog, and reserve the Claude catalog for work where its output quality earns the difference.

Getting started and honest limits

Signup is a username, a password, and an email verification code — about a minute — and new accounts receive ¥1 of credit, which is enough to run every step in this guide. Top-ups are Alipay only, starting at $1, with ¥1 = $1 of quota; invites give ¥1 to both sides. Two limits worth stating plainly: there is no card or PayPal payment yet, so this suits users who can pay with Alipay; and it is a small, new service, so treat the free credit as a way to verify your own workflow before committing real volume. Questions and failure reports go to yangda611@gmail.com.

Disclosure: I operate ZeroShell. Every price and behavior above is from the public pricing page and documented, tested endpoint behavior as of November 2026; check the pricing page for current figures.