Posted on Sep 30

Running Claude Code and Codex CLI Behind a Single OpenAI-Compatible Endpoint

#opensource #tooling #api #ai

If you are using AI coding agents like Claude Code or Codex CLI and you keep hitting the same three walls - regional access, overseas card requirements, and one API key per model - this post is a practical writeup of how I consolidated everything into a single endpoint.

The problem

My setup used to look like this:

Claude Code -> Anthropic key (needs an overseas card)
Codex CLI -> OpenAI key (needs an overseas card)
Local tools (Cherry Studio / LobeChat / NextChat) -> each with its own key

Three wallets, three billing dashboards, three points of failure.

The approach: one base_url, both protocols

The cleanest fix is a gateway that speaks both the OpenAI and the Anthropic protocol on the same host, so every tool points at one base_url.

export OPENAI_BASE_URL="https://api.dshapi.icu/v1"
export OPENAI_API_KEY="sk-..."

For Claude Code:

export ANTHROPIC_BASE_URL="https://api.dshapi.icu"
export ANTHROPIC_AUTH_TOKEN="sk-..."
export ANTHROPIC_MODEL="claude-sonnet-4-5"

That is the whole config. No per-tool proxy, no key juggling.

Endpoint verification (I actually ran these)

Endpoint	Protocol	Result	Latency
GET /v1/models	OpenAI	200	1.61s
POST /v1/chat/completions	OpenAI	200	2.19s
POST /v1/responses	OpenAI Responses	200	2.37s
POST /v1/messages	Anthropic	200	3.94s

Requesting /v1/models without a key correctly returns 401, so auth is enforced.

Quick smoke test:

curl https://api.dshapi.icu/v1/models -H "Authorization: Bearer $OPENAI_API_KEY"

curl https://api.dshapi.icu/v1/chat/completions \

  -H "Authorization: Bearer $OPENAI_API_KEY" \

  -H "Content-Type: application/json" \

  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"hi"}]}'

Cost: where the savings actually come from

Cost numbers without a mechanism are just marketing. Here is the mechanism, taken from a real call:

Metric	Value
prompt_tokens	1,755
prompt_cache_hit_tokens	1,536 (87.5%)
prompt_cache_miss_tokens	219
completion_tokens	20

Two things compound:

Prompt cache hits - 87.5% of input tokens were served from cache, which bills far below normal input.
Group multiplier - the open-weights group bills at 0.08x of list price.

Aggregated over 30,734 requests and 4.07B tokens: 28.67 CNY paid against 359.51 CNY at list price, i.e. 7.97%.

Models available

deepseek-v4-flash      deepseek-v4.1-flash    deepseek-v4-pro

glm-5.2                glm-5.3                glm-5.3-flash

kimi-k2.8              kimi-k3                minimax-m3

hy3                    hy4

Registration friction

Item	Requirement
Email	Any (QQ mail works)
Overseas credit card	Not required
Payment	WeChat Pay / Alipay
Billing	Pay-as-you-go, balance does not expire

Caveats

This is an aggregator, not the official vendor. If you need enterprise SLAs or data-residency guarantees, use the official APIs.
I am on the referral program, so the last link pays me roughly 10%. The plain domain works identically if you would rather avoid that.
Latency is region dependent; these measurements come from a residential connection in Asia.

DEV Community