Two environment variables and Codex CLI runs against our OpenAI Responses-compatible endpoint. No subscription, no card, no minimum: prepaid credits are consumed per token at 30% of the official list price, and the line stops the moment the balance reaches zero.
macOS / Linux
export OPENAI_BASE_URL=https://api.mddtech.cn/p/v1 export OPENAI_API_KEY=sk-your-key codex
Windows PowerShell
$env:OPENAI_BASE_URL="https://api.mddtech.cn/p/v1" $env:OPENAI_API_KEY="sk-your-key" codex
| What you pay | Official-priced usage you get | Typical use |
|---|---|---|
| $5 | ≈ $16.67 | a few evenings of agent work |
| $10 | ≈ $33.33 | a light week of Codex CLI |
| $20 | ≈ $66.67 | heavy use, multiple repos |
Rates follow the official published token prices × 0.30. Credits do not expire and the balance cannot go negative.
Only models that genuinely work are enabled — a request for an unsupported model is rejected before it reaches upstream and is not billed. We would rather expose three working models than twenty that fail halfway through a task.
Yes. Codex CLI honours OPENAI_BASE_URL and OPENAI_API_KEY, so any endpoint that
implements the OpenAI Responses API works. Ours does, and it is Anthropic-compatible too, so the same portal
serves both tools.
Upstream capacity is third-party, so it can degrade — that is the honest risk of a relay. What we control is the model list (kept in sync with what actually works) and transparency: the live status panel on the home page shows real probe latency and availability, not an invented uptime figure.
USDC on Base (recommended) or USDT on Polygon / BNB Chain, sent to a one-time address with a unique amount. Matching is automatic and credit usually lands within a minute — no card details ever reach this service.
Yes, self-service and without an email address, so you can benchmark the endpoint from your own machine before paying.
No. This is an independent third-party relay. Model names describe API compatibility only.