Codex CLI through PromptGate

Codex CLI uses OpenAI’s newer Responses API (POST /v1/responses) — different envelope from Chat Completions, agentic-by-default with multi-step tool calls. PromptGate’s Agent Proxy speaks that surface natively, so Codex routing is the same one-env-var move.

Prerequisites

An Agent Proxy project (see the Claude Code recipe for setup).
An OpenAI credential under Credentials, OpenAI provider toggled on under the project’s Providers entry.
An API token with the chat scope.

Setup

export OPENAI_BASE_URL="https://your-promptgate-host/api/<project-uuid>/v1"
export OPENAI_API_KEY="pg_live_..."

The /v1 suffix is important — Codex appends /responses to whatever you set as the base URL.

Restart Codex CLI. Run any session.

What you get on the Responses API surface

The exact same observability as on Chat Completions:

Live Logs show every request with the input + output items + tool calls
Cost Dashboard aggregates spend across Codex AND any other agent in the same project
Reversible Redaction runs on the input field and instructions field — works the same as on Chat Completions
Secret Scanner intercepts before the request reaches OpenAI

Cross-routing trick

Because PromptGate translates between API shapes internally, you can target a non-OpenAI provider from Codex:

# In your wrapper aliases, define:
#   alias "smart" → anthropic:claude-3-5-sonnet
#
# Then Codex calls /v1/responses with model: "smart" and gets
# a Claude response wrapped in OpenAI's Responses envelope.

This is useful when:

Anthropic gives a better response on a particular task but you’ve standardized on the Codex CLI ergonomics.
You want to A/B different providers without changing the agent’s UX.

Troubleshooting

Cannot find route POST /api/<uuid>/v1/responses: the project type is wrong. Codex expects /v1/responses on the Agent Proxy or AI Wrapper project type.
400 with invalid_project_type: the token belongs to an ai_gateway project — those use the per-endpoint URL /api/<uuid>/<slug>, not the wrapper-style URL.

Cost note

Codex CLI defaults to OpenAI’s frontier model. Set up a routing rule that drops simple-tool-result-summarization calls to gpt-4o-mini. Typical agent-style workloads have a long tail of small post-tool-call summaries that don’t need a frontier model — easy 5-10× cost reduction.

© Akyros Labs LLC. All rights reserved.