How to Use Kimi with Codex CLI (Step-by-Step Setup Guide)

Wed, 19 Aug 2026 16:59:13 +0800

Codex uses OpenAI models by default, but it also supports third-party model providers. Kimi Code exposes an endpoint compatible with the Codex Responses API. This guide documents how to use Kimi K3 as a Codex model.

Prerequisites

Install Codex CLI: npm i -g @openai/codex
Open the Kimi Code console through https://www.kimi.com/code and create an API key
Make sure Codex can run in the current terminal and that the terminal can reach https://api.kimi.com/coding/v1

Configure the Provider

Codex’s user-level configuration directory is ~/.codex. Create kimi.config.toml and configure Kimi Code as a model provider:

model = "k3"
model_provider = "kimi"
model_context_window = 1048576
model_supports_reasoning_summaries = false
web_search = "disabled"

[model_providers.kimi]
name = "Kimi Code"
base_url = "https://api.kimi.com/coding/v1"
env_key = "KIMI_API_KEY"
wire_api = "responses"
stream_idle_timeout_ms = 600000

The configuration filename and profile name correspond to each other. When kimi.config.toml is placed under $CODEX_HOME, Codex can load it with --profile kimi. If CODEX_HOME is not set, the default directory is ~/.codex.

Here is what the key settings do:

model = "k3": Uses the Kimi K3 model.
model_provider = "kimi": References the [model_providers.kimi] section below.
model_context_window = 1048576: Declares the model’s context window, which is useful for large codebases and long tasks.
model_supports_reasoning_summaries = false: Disables the reasoning summaries capability declaration so the client does not process the response using an incompatible format.
web_search = "disabled": Disables Codex web search for this profile; requests use the Kimi coding endpoint only.
base_url: The Kimi Code Responses API endpoint.
env_key: Tells Codex which environment variable contains the API key.
wire_api = "responses": Uses the Responses API protocol.
stream_idle_timeout_ms = 600000: Sets the streaming idle timeout to 600,000 milliseconds, or 10 minutes, allowing more time for long coding tasks.

Configure the API Key

Add the following to ~/.zshrc or ~/.bashrc:

export KIMI_API_KEY="your-kimi-api-key"

Then restart the terminal, or run the following command to apply the change:

source ~/.zshrc

If you use Bash, run source ~/.bashrc instead. The API key is passed through the environment variable, so do not put it directly in a configuration file or alias that will be committed to Git.

Start Codex with Kimi

After completing the setup, start the Kimi version of Codex with the --profile option:

codex --profile kimi

Codex will load $CODEX_HOME/kimi.config.toml and send Kimi K3 requests to https://api.kimi.com/coding/v1. To use the default OpenAI model, start Codex without a profile:

codex

Switch Quickly with a Shell Alias

Typing --profile kimi every time can be tedious. Define an alias in your shell configuration instead:

alias codexkimi="codex --profile kimi"

You can then enter codexkimi to start Codex with Kimi, or codex to use the default model, without repeatedly editing the TOML configuration.

My Experience

The important part of this setup is not changing Codex itself, but using its model provider mechanism to route requests to Kimi Code. Kimi K3’s one-million-token context window is useful for large codebases and long-running tasks. With stream_idle_timeout_ms set to 10 minutes, long tasks also have more time before an idle streaming response is treated as a timeout.

Keep in mind that web_search = "disabled" means Codex will not use web search. If a task depends on up-to-date web information, prepare the relevant material separately or use another search tool outside this profile. API usage is billed to the Kimi Code account, so check the account balance and service terms before using it.

Final Thoughts

Codex’s model provider configuration makes switching to Kimi Code straightforward: create an API key, expose it through KIMI_API_KEY, then start Codex with kimi.config.toml and --profile kimi. For workflows that need a large context window and a Kimi coding model, this is a convenient way to connect the two tools.

Kimi | Attack on Life