Use PaletteAI Inference Launchpad with OpenAI Codex

This guide explains how to connect the OpenAI Codex CLI to a PaletteAI Inference Launchpad appliance so that a model running on the appliance serves every request instead of OpenAI's hosted API. Codex connects to the appliance the same way it connects to an OpenAI endpoint, so you trust the appliance certificate, add a custom model provider to the Codex configuration file, and confirm the connection.

The console assembles these values for you. In the Connect a coding agent panel, the Codex CLI tab lists the steps and generates the configuration for bash, zsh, or PowerShell. For what each field does, refer to OpenAI Codex Configuration.

Prerequisites

• The OpenAI Codex CLI installed and already working. For installation, refer to the OpenAI Codex repository.
• A running PaletteAI Inference Launchpad appliance with at least one model deployed and serving. To deploy a model, refer to Deploy a Model.
• An API token for the appliance. To create one, refer to Generate an API Token, or use a token an administrator generated for you.
• Network access from your machine to the appliance. Codex sends each request from the machine it runs on, so an appliance on a private network works.

Download the Platform CA Certificate

When the appliance serves HTTPS with a certificate issued by its own platform Certificate Authority (CA), the connect panel offers that CA certificate as a download. Your machine does not trust the platform CA yet, so download the certificate before you configure the agent.

1. In the console, open the Connect a coding agent panel. Select Connect Coding Agent on the homepage, select Connect a Coding Agent in the Create API token dialog that reveals a new token, or select Connect Agent in the dialog that finishes the Add client wizard.

2. Select the tab for your agent, and then select CA certificate. The browser saves palette-ai-inference-launchpad-ca.crt to your Downloads folder.

If the panel shows no CA certificate step, the appliance does not need one. Skip this section and every later step that points at the downloaded file.

Configure Codex

1. Add the following custom provider to the Codex configuration file at ~/.codex/config.toml. Use the home folder, not a project folder. Replace <appliance-host> with your appliance address.

   model = "gpt-5.6"
   model_provider = "launchpad"
   
   [model_providers.launchpad]
   name = "PaletteAI Inference Launchpad"
   base_url = "https://<appliance-host>/v1"
   wire_api = "responses"
   env_key = "LAUNCHPAD_API_KEY"

2. Run the following commands in your shell. Codex fails to parse an export line inside config.toml, so both values stay in the shell. Replace <lpai-token> with your token.

   export CODEX_CA_CERTIFICATE=$HOME/Downloads/palette-ai-inference-launchpad-ca.crt
   export LAUNCHPAD_API_KEY=<lpai-token>

   Include the CODEX_CA_CERTIFICATE line only when the panel shows the CA certificate step. Codex layers that certificate on top of the system trust store, so certificate verification stays on. On PowerShell, the panel writes the same path as $env:USERPROFILE\Downloads\palette-ai-inference-launchpad-ca.crt.

Verify the Connection

1. Run a single prompt to confirm the appliance answers. The --skip-git-repo-check flag lets you run the test outside a git repository.

   codex exec --skip-git-repo-check "reply with exactly CODEX_OK and nothing else"
   Expected output

   CODEX_OK

   A reply confirms that the certificate, base URL, token, provider, and model routing all work.

2. Start an interactive session.

   codex

info

Codex may print a Model metadata for <model> not found warning. This warning is cosmetic and does not affect the request.

Request Routing and Quotas

Requests you send through the appliance are subject to the routing rule and quota configured for your client. If an operator configured a routing rule for your token, the appliance can redirect a request to a frontier model instead of a local one. Quotas apply to the client, so every API token that belongs to the client draws on the same limits. When the client exhausts a quota, the appliance returns an HTTP 429 response.

To understand what a client is and how API keys and quotas govern usage, refer to Clients and Quotas. To understand how the Tier map and the semantic router combine to pick a model for each request, refer to Routing Behavior.

Next Steps

To look up each configuration value, refer to OpenAI Codex Configuration. To connect a different coding tool, refer to Use PaletteAI Inference Launchpad with Claude Code or Use PaletteAI Inference Launchpad with Cursor.