Skip to main content
Codex CLI talks to models over the OpenAI Responses API. The gateway serves POST /v1/responses, so pointing Codex at it is a provider entry in Codex’s config.toml. No Tracelane SDK is involved.

​
Setup

Mint a key with the chat scope in Settings → API keys. Add the key for the model’s provider under Settings → LLM providers, because the gateway calls that provider with your own key. Then add this to ~/.codex/config.toml:
Codex sends Authorization: Bearer $TRACELANE_API_KEY to {base_url}/responses. Each call becomes one gateway span, with the same keys, budgets, rate limits and guardrails as every other route.

​
Using a model that is not OpenAI’s

Set model to any model the gateway routes, for example a claude-* or gemini-* model:
How the call reaches the provider depends on the provider: Translated calls have these limits:
  • No stored state. previous_response_id, background and conversation return 400 unsupported_parameter. Codex with store = false, its default, sends the full conversation on every turn and is not affected.
  • Hosted search is dropped, not run. Codex’s web_search and tool_search tools are removed before the call, and the response header x-tracelane-dropped-tools lists them. Other hosted tools, such as code_interpreter, file_search and computer_use, return 400.
  • Reasoning items are not returned. Encrypted reasoning from an earlier OpenAI turn means nothing to another provider, so the gateway drops it and counts it on the span.
  • Reasoning effort, structured output and file inputs are translated into the provider’s own controls where it has them, for example adaptive thinking on Claude or thinkingLevel on Gemini. A value the provider cannot honour returns 400 unsupported_parameter naming the field.
  • Two hints are dropped, and the drop is reported. text.verbosity has no equivalent outside OpenAI. parallel_tool_calls: false is dropped on providers that have no such switch, so the model may return more than one call in a turn. Both appear in x-tracelane-dropped-tools. Nothing is ignored silently.
Codex’s freeform apply_patch tool is passed to the provider as a tool that takes a single string. The call comes back to Codex as a custom_tool_call, the shape Codex expects.

​
Errors

Gateway errors use the OpenAI shape, {"error": {"message", "type", "code", "param"}}. For a relayed provider, an upstream error keeps its status and body, with credentials scrubbed. A rejected provider key is reported as provider_key_rejected and the upstream body is not returned.