Use Kimchi Router in Codex | Kimchi Inference

This guide shows you how to run Codex, OpenAI's coding agent, on the Kimchi Model Router. Tested with Codex 0.159.0.

Once connected, Codex sends its requests to the auto model, and the Model Router picks the model that fits each task.

Prerequisites

The Kimchi CLI installed and set up. See Kimchi CLI.
Codex 0.159.0 or later.

Connect Codex to Kimchi

Run kimchi setup-tools

Run the setup wizard and select Codex:

kimchi setup-tools

The wizard configures Codex to send its requests to the Kimchi gateway.

Start a session on auto

Start Codex with the auto model:

codex --model auto

To switch a running session, select auto in the /model dropdown.

Verify the connection

Ask Codex a simple question, then open the Kimchi console. Go to Analytics and set the time range to 24h. Your requests appear in the totals within seconds.

To see which models the Router picked, see Router analytics.

Supported features

Terminal sessions
The ChatGPT desktop app
Image support

Limitations and known issues

Context window

Codex shows no context window indicator for the auto model. Users report an effective window of about 256K tokens.

You can set model_context_window in ~/.codex/config.toml to plan for a larger window. Treat this override with care: Codex has known issues where custom context window values break auto-compaction.

Kimi K3 request failures

Some requests fail when the Model Router picks Kimi K3. This is a known issue on the Kimchi gateway side, and it affects Codex sessions that use the Responses API.

Next steps

Kimchi Model Router

How routing works, pricing, and Router analytics.

Available models

The serverless model catalog the Router picks from.