Run open models in Codex

Codex runs on Keln in the ChatGPT desktop app and in the codex command line. Pointed at Keln, it lists the open-source models in its own model picker, each with effort levels. Your ChatGPT sign-in is untouched, and you can switch back whenever you want.

What you need to start

The ChatGPT desktop app for macOS.
A Keln API key. Sign up at app.keln.ai, then open Keys → Create key. Keln shows the key once.
Credit on your Keln account. Billing → Add credit.

Connect Codex to Keln

1. Download the model catalog

Download keln-codex-models.json and save it as ~/.codex/keln-models.json. Codex builds its model picker and effort control from this file. It lists the Keln models and Codex's agent instructions, and holds no secrets.

2. Point Codex at Keln

Open ~/.codex/config.toml (create it if it does not exist). Add these four lines at the very top of the file:

# Keln
model = "zai-org/glm-5.3"
model_provider = "keln"
model_catalog_json = "/Users/YOU/.codex/keln-models.json"   # full path; ~ is not expanded
model_reasoning_effort = "high"

And add this block at the very end of the file:

[model_providers.keln]
name = "Keln"
base_url = "https://api.keln.ai/v1"
experimental_bearer_token = "sk-keln-…"   # your key

Replace YOU with your macOS user name, and sk-keln-… with your Keln API key. The app does not read shell variables, so the key has to live in the file. Keep the file private:

chmod 600 ~/.codex/config.toml

3. Reopen ChatGPT

Quit and reopen the app. Codex reads both files at startup. Open Codex from the sidebar. The message box shows the model and effort, and the account menu at the bottom of the sidebar reads Keln.

Codex in the ChatGPT app: the message box shows the current model and effort, and the account chip at the bottom of the sidebar reads Keln

The account menu open, with Keln as its header

4. Pick a model and effort

Click the model chip in the message box. The slider sets the effort; click the model name for the list.

The effort slider above the model chip, set to High for Kimi K3

The model picker listing GLM 5.3, GLM 5.3 Flash, Kimi K3, DeepSeek V4 Flash and DeepSeek V4 Pro

Models and features

Model	Role	Input
GLM 5.3	Frontier	text
Kimi K3	Frontier	text + images
DeepSeek V4 Pro	Frontier	text
GLM 5.3 Flash	Flash	text + images
DeepSeek V4 Flash	Flash	text

Effort levels. Low, High and Extra High, from the slider in the message box. DeepSeek V4 Pro has High and Extra High only.
Web search. Codex's search runs on Keln, server side, priced per search (pricing, details).
Zero data retention, the same privacy contract as the API (Privacy & zero data retention).

Keln adds models over time. To ask for one, open Model catalog in the portal and send a model request.

Codex on the command line

The same config.toml drives the codex command. On the command line the key can come from your shell instead of the file. Replace the experimental_bearer_token line with env_key = "KELN_API_KEY", then run export KELN_API_KEY=sk-keln-….

Switch between Keln and ChatGPT

To go back to ChatGPT, comment out the four lines at the top of config.toml (model, model_provider, model_catalog_json, model_reasoning_effort) and reopen the app. Uncomment them to return to Keln.

Using Claude Desktop instead? See Run open models in Claude Desktop.