Run Codex CLI on it
Codex speaks OpenAI's Responses API and nothing else; this service serves it at /v1/responses with no shim.
https://189.24.66.33/v1/responsesThe environment
export OPENAI_BASE_URL=https://189.24.66.33/v1
export OPENAI_API_KEY=sk-relay-...
codex --model glm-5.3-flash-xplodedThe base URL carries /v1 here: the Responses client appends only /responses itself.
Forwarding and metering
Responses requests share the same metering pipeline as every other format: holds, settlement, receipts and logging behave identically. In streaming mode the real Responses event sequence is forwarded as-is — response.created, output items, text deltas and response.completed arrive in order.
What is not available
previous_response_id is refused with a clear error: continuing a stored response would mean the server kept your conversation, and it keeps no prompts at all. Codex sets store: false and sends the full input each turn, so nothing is lost. Hosted tools (web_search, file_search, computer_use) run outside the model and are unavailable; function tools you define work normally.
The model must call tools
An agent is useless against a model with no tool support. Every model in the catalogue calls tools, so point it at glm-5.3-flash-xploded with no driver model to configure. See tool calling.