Skip to main content

OpenAI Codex + MLflow AI Gateway

Route OpenAI Codex through the MLflow AI Gateway to get centralized tracing and observability, while each developer authenticates with their own OpenAI subscription.

Prerequisites​

  • MLflow server running with a SQL backend (mlflow server --port 5000)
  • Codex installed (npm install -g @openai/codex)

Step 1: Create an OpenAI Endpoint​

Navigate to the AI Gateway tab at http://localhost:5000/#/gateway and click Create Endpoint.

  • Provider: OpenAI
  • Model: choose any model as the actual model is selected by Codex CLI
  • Endpoint name: choose a name, e.g. my-codex-endpoint
  • LLM Connection: select an existing connection or create a new one (see Create an LLM Connection)
tip

The server-side API key in the LLM Connection can be set to a dummy value (e.g. dummy). The gateway detects Codex's User-Agent and forwards the client's own credentials.

Step 2: Run Codex​

Configure Codex to point to the gateway base URL and run it.

bash
codex --config 'openai_base_url="http://localhost:5000/gateway/proxy/my-codex-endpoint/v1"'

For a persistent setup, add the same value to ~/.codex/config.toml:

~/.codex/config.toml
toml
openai_base_url = "http://localhost:5000/gateway/openai/v1"

Note that you need to authenticate with your API key instead of ChatGPT subscription.

What You Get​

Every session is captured as an MLflow trace. Open the Logs tab in the MLflow UI to inspect inputs, outputs, token usage, and latency for every request.

Codex trace in MLflow

Usage Tracking

Monitor token usage and costs across all Codex sessions

View usage tracking →

Guardrails

Add content policies to all Codex requests automatically

Configure guardrails →

Budget Alerts & Limits

Set spending limits globally or per workspace to keep sessions within budget

Configure budget limits →