Use OminiGate in OpenAI Codex

Add a custom model_provider in config.toml and Codex CLI will route every request through OminiGate — OpenAI, Anthropic, and Gemini models included.

Codex

Overview

OpenAI's Codex CLI reads provider definitions from ~/.codex/config.toml. Each provider has a name, a base_url, an auth env var, and a wire protocol. That means Codex can call any OpenAI-compatible gateway — including OminiGate — without a patch or fork.

For OminiGate, pick wire_api = "chat" (the OpenAI chat-completions format). Codex will send requests to /v1/chat/completions on your base URL and attach the OMINIGATE_API_KEY header for you.

  • One config file, zero code changes.
  • Switch between OminiGate and OpenAI with a single flag.
  • Works with every model OminiGate exposes on the chat-completions endpoint.

Prerequisites

You'll need Codex CLI installed and an OminiGate API key.

1. Install Codex CLI

Codex ships as an npm package (@openai/codex). Node 18+ is required. macOS and Linux are supported natively; Windows runs through WSL2.

npm i -g @openai/codex

2. Get an OminiGate API key

Create a key in the dashboard. The key will be referenced via the OMINIGATE_API_KEY environment variable, not stored in config.toml.

Open API keys→

Configure Codex CLI

Codex loads configuration at startup from ~/.codex/config.toml. We'll register OminiGate as a custom provider and wrap it in a profile — profiles let OminiGate coexist with Codex's built-in OpenAI usage instead of replacing it.

Clear existing OpenAI environment variables first

A lingering OPENAI_API_KEY or OPENAI_BASE_URL in your shell can clash with the Codex profile, especially when switching between OminiGate and vanilla OpenAI. Unset them before moving on.

shell
unset OPENAI_API_KEY
unset OPENAI_BASE_URL

1. Register provider + profile

Append the block below to ~/.codex/config.toml (create the file if it doesn't exist). The [profiles.ominigate] section means your existing Codex workflow stays intact — activate this profile only when you want to hit OminiGate.

~/.codex/config.toml
# ~/.codex/config.toml

[model_providers.ominigate]
name = "OminiGate"
base_url = "https://api.ominigate.ai/v1"
env_key = "OMINIGATE_API_KEY"
wire_api = "chat"
requires_openai_auth = false
request_max_retries = 4
stream_max_retries = 10
stream_idle_timeout_ms = 300000

[profiles.ominigate]
model = "openai/gpt-5.4-pro"
model_provider = "ominigate"

Why each field matters: wire_api="chat" tells Codex to use OpenAI chat-completions (not the Responses API). requires_openai_auth=false disables the default OpenAI OAuth flow, which doesn't apply to third-party providers. request_max_retries / stream_max_retries / stream_idle_timeout_ms are Codex's production-grade retry and long-stream timeouts — copy these values unless you have a reason not to.

2. Export your API key

Codex reads the key from the env var you named in env_key. Add this to your shell profile so it survives new sessions.

export OMINIGATE_API_KEY="sk-omg-your-api-key"

Heads up: OpenAI has announced that wire_api = "chat" is deprecated and will be removed from Codex CLI in 2026. OminiGate currently exposes /v1/chat/completions only; we're evaluating a Responses API endpoint so Codex keeps working after the deprecation. Watch the changelog. For now, "chat" still works and you may see a deprecation warning on startup.

Run Codex

Activate the profile you just defined. Codex loads the OminiGate model_provider plus the default model slug, then starts the session:

shell
# Use the profile you defined in config.toml
codex --profile ominigate "Refactor the main.rs to use async/await"

Override the default model for one invocation:

shell
codex --profile ominigate --model anthropic/claude-opus-4.6 "Explain this repo"

Running plain codex (no --profile) still uses Codex's built-in OpenAI setup — the OminiGate profile is opt-in per command.

Recommended models

Codex uses the OpenAI chat-completions format, so any model on OminiGate that supports that protocol works. These four are good starting points.

openai/gpt-5.4-pro

Top-tier OpenAI reasoning model. Best for complex refactors and multi-file edits.

openai/gpt-5.3-codex

OpenAI's Codex-tuned variant. Optimized for code generation, multi-file edits, and tool-use loops — the natural default when running Codex CLI.

anthropic/claude-opus-4.6

Anthropic's highest tier, also exposed on the chat-completions endpoint. Strong at long-context code review.

google/gemini-3.1-pro-preview

Google's flagship preview model — large context window, multimodal inputs.

Browse the full catalog at /models.

Troubleshooting

Codex keeps calling api.openai.com instead of OminiGate.

Check that model_provider = "ominigate" is set at the top level of config.toml, not inside a [profiles.*] block. Run codex --verbose [prompt] to see which provider Codex resolved.

I get 404 Not Found on every request.

The most common cause is wire_api = "responses". OminiGate exposes /v1/chat/completions, not the OpenAI Responses API. Switch to wire_api = "chat" and restart.

Codex complains that OMINIGATE_API_KEY is missing.

Codex reads env vars from the shell that launched it. If you set the var in a new terminal, relaunch Codex from that shell — or add the export line to your ~/.zshrc / ~/.bashrc.

Next steps

Explore the full API reference or browse more models you can call from Codex.

API referenceBrowse models