Codex CLI

OpenAI Codex CLI is OpenAI's official terminal coding agent, built on the Responses API. With WRouter you can drive Codex with one key that also reaches Claude, Gemini, and other non-OpenAI models.

Prerequisites

Item	Requirement
Codex CLI	≥ 0.5 (binary codex)
Node.js	≥ 20 (when installing via npm)
WRouter	A working token

Install

npm i -g @openai/codex

# Verify
codex --version

Configure

1. ~/.codex/config.toml

model_provider = "wrouter"
model = "gpt-5.5"

[model_providers.wrouter]
name = "wrouter"
base_url = "https://wrouter.ai/v1"
wire_api = "responses"
env_key = "WROUTER_API_KEY"

Field reference:

Field	Meaning
base_url	WRouter OpenAI-compatible base. Must include /v1
wire_api	"responses" (recommended, Agentic) or "chat" (Chat Completions)
env_key	Env var to read the API key from

2. Environment

# macOS / Linux
export WROUTER_API_KEY="sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

# Windows (PowerShell)
[Environment]::SetEnvironmentVariable("WROUTER_API_KEY", "sk-xxxx...", "User")

Verify

codex

Run /status:

Provider:    wrouter
Model:       gpt-5.5
Endpoint:    https://wrouter.ai/v1
Wire API:    responses

Smoke test:

> What files are in the current directory?

Codex should invoke the shell tool.

Recommended models

Model	wire_api	When
gpt-5.5	responses	Primary; best native Agentic support
o3	responses	Reasoning-heavy
gpt-5.4	responses or chat	General purpose
claude-sonnet-4-6	chat ⚠️	Stick to chat for Claude
gemini-2.5-pro	chat ⚠️	Same

Switch in-session:

/model claude-sonnet-4-6

⚠️ Important: when switching to non-OpenAI models, also flip wire_api to "chat", otherwise multi-step tool loops may degrade. Best practice: declare multiple providers:

[model_providers.wrouter-chat]
name = "wrouter (chat)"
base_url = "https://wrouter.ai/v1"
wire_api = "chat"
env_key = "WROUTER_API_KEY"

# Then: codex --provider wrouter-chat

Troubleshooting

Symptom	Likely cause	Fix
401 Unauthorized	WROUTER_API_KEY missing/wrong	echo $WROUTER_API_KEY; ensure toml env_key matches
404 / unsupported wire_api	Model doesn't support Responses	Set wire_api = "chat" or switch model
Empty / hanging responses	wire_api mismatched with model	Same
Garbled / tool failures	Codex too old	npm i -g @openai/codex@latest
rate_limit_exceeded	Hit WRouter RPM cap	Adjust in console

Known limitations

• Some Codex high-level features (e.g. the built-in web_search tool) rely on OpenAI-only upstream paths and aren't available when routing non-OpenAI models through WRouter.
• wire_api = "responses" only works when the model itself implements Responses — today primarily OpenAI's GPT-5.x / o-series family.