Codex CLI with UberLLM: run OpenAI’s coding agent on any model

Steps

1. Make a key on your keys page and add credit.
2. Install Codex if you do not have it: npm install -g @openai/codex.
3. Put the key in your shell profile (~/.zshrc or ~/.bashrc) and open a new terminal:

export UBERLLM_API_KEY="ull_..."

4. Open ~/.codex/config.toml (create it if it is missing) and add:

model = "moonshotai/kimi-k3"
model_provider = "uberllm"
web_search = "disabled"

[model_providers.uberllm]
name = "UberLLM"
base_url = "https://api.uberllm.dev/v1"
env_key = "UBERLLM_API_KEY"
wire_api = "responses"

5. Run codex in a project folder and give it a task.

Keep base_url ending in /v1: Codex adds /responses to it. env_key is the name of the variable, not the key itself.

Switch models

Pass -m for one session, for example codex -m deepseek/deepseek-v4.1-flash for a small fix, or change model in the file. Any chat model id from uberllm.dev/models works.

For scripts and CI, codex exec runs one task without the interactive screen:

codex exec --skip-git-repo-check "add a docstring to every function in utils.py"

Settings worth adding

• model_context_window: Codex does not know the context size of models outside OpenAI's list and assumes 272,000 tokens. For a smaller model, set the number from its page on uberllm.dev/models (for example model_context_window = 131072), so Codex compacts the conversation before it runs out of room.
• model_reasoning_effort: "low", "medium" or "high" is passed on to models that think (DeepSeek, Kimi Thinking, GLM, GPT-5). Leave it out for the model's default.
• web_search = "disabled": Codex's web search runs on OpenAI's own servers, so it cannot run through another provider. With it off, the model is not offered a tool it cannot use.

What works

Every Codex tool is a normal function call from the API's side, so command runs, file edits, plans and MCP servers you add to Codex all work on any model that takes tool calls. Codex sends the whole conversation on every turn (store = false), so nothing is kept on UberLLM between calls. Each call shows in Dashboard → Activity under the app name codex, with its cost.

Troubleshooting

• wire_api = "chat" is no longer supported: Codex 0.95 and later only speak the Responses API. Set wire_api = "responses".
• 401 or "Missing environment variable: UBERLLM_API_KEY": the variable is not exported in the terminal Codex runs in. Open a new terminal after editing your profile, or run export UBERLLM_API_KEY=... first.
• 404 on every turn: base_url is missing /v1. It must be exactly https://api.uberllm.dev/v1.
• "Model metadata for … not found. Defaulting to fallback metadata": Codex prints this for every model outside OpenAI's own list. It is harmless; set model_context_window as above.
• The model answers but never runs a command: pick a model from the first table. Small models drift out of Codex's tool format.
• 402 insufficient credits: add credit on the credits page. For a hard ceiling, set a spend limit on the key.

Keep reading

• Codex doc · Responses API reference
• OpenCode setup · Cline setup