1. Step 1: Copy and save the config

    routerhub.config.toml
    model_provider = "routerhub"

[model_providers.routerhub]
name = "RouterHub"
base_url = "https://routerhub.si/v1"
env_key = "ROUTERHUB_API_KEY"
wire_api = "responses"
stream_idle_timeout_ms = 600000

    macOS / Linux

    Save as ~/.codex/routerhub.config.toml.

    Windows

    Save as %USERPROFILE%/.codex/routerhub.config.toml.

    Create any missing folders. If the file already exists, merge these settings without replacing your other settings. Keep your key private and out of repositories.

    Download config. Your usual Codex config and saved login stay unchanged.

  2. Step 2: Set your API key

    Replace the placeholder with your RouterHub key and run:

    macOS / Linux

    macOS / Linux
    export ROUTERHUB_API_KEY='PASTE_ROUTERHUB_KEY_HERE'

    Windows (PowerShell)

    Windows (PowerShell)
    $env:ROUTERHUB_API_KEY = 'PASTE_ROUTERHUB_KEY_HERE'

    This sets the key for this terminal. Keep it and its command history private; close it when finished.

  3. Step 3: Start Codex

    Copy a model ID with Responses support from the models and pricing page. Replace MODEL_ID below and run in the same terminal:

    Run in your terminal
    codex --profile routerhub --model MODEL_ID

    Check that Codex shows RouterHub and your chosen model. To switch, quit and launch with another Responses model ID.

Need to install Codex CLI?

Install a current Codex CLI with npm install -g @openai/codex, or use the official installer. This profile format needs Codex 0.134.0+.

Check the connection (optional)

When you’re ready to use credits, open an empty project and send this short test:

Reply with OK. Do not read files, run commands, or make edits.

Open RouterHub usage and check that the request appears under the right model. Keep auto-approval off while testing. A text reply checks the connection, not all coding features.

Troubleshooting

A stream stops while the model is thinking

The provider sets stream_idle_timeout_ms = 600000: ten minutes without an SSE event, not a whole-request or tool-execution limit. Codex does not document OpenCode-style header/chunk/total timeout keys in this provider config, so do not add them. This does not disable the client’s built-in retry behavior. Check Usage and the request ID before manually retrying.

Codex still shows OpenAI or ChatGPT

Launch with --profile routerhub. On Codex 0.134.0 or later, use routerhub.config.toml beside your user config, not an old [profiles.routerhub] table. Inspect startup/status information before sending a prompt.

Missing key or 401

Enter a RouterHub key in the same terminal. The profile reads ROUTERHUB_API_KEY with env_key; do not replace or delete your saved ChatGPT credentials.

404 or model unavailable

Use https://routerhub.si/v1 and an exact Responses-capable ID. Do not append /responses to the base URL or substitute a Chat Completions-only model.

Streaming, reasoning, or tools fail

Stop repeated retries. Send support the redacted error, model ID, CLI version, and request ID. This setup has not been tested against RouterHub with authenticated end-to-end requests.

Still stuck? Contact support.

Common questions

Does this use my ChatGPT subscription?

No. The RouterHub profile uses your RouterHub API key and prepaid balance. Your saved ChatGPT login is left alone.

Can Codex use every RouterHub model?

Not through this setup. It uses wire_api = "responses", so choose a model with Responses support on the models page. That is not a guarantee of every Codex feature.