Kimi K3’s direct Moonshot API uses Chat Completions, while current Codex custom providers support the Responses wire API. Requesty translates between the two formats, so Codex can use Kimi K3 without an obsolete wire_api = "chat" setting.
Kimi K3 is Moonshot’s current flagship model, with native vision, up to a 1-million-token context window, and configurable reasoning effort. Feature support inside Codex depends on what the Chat Completions compatibility layer carries through.
Install Codex and create a key
Follow the current Codex Windows and WSL setup, then create a Moonshot API key.
Open ~/.codex/config.toml and merge in the following provider:
model = "moonshot/kimi-k3"
model_provider = "requesty"
model_supports_reasoning_summaries = false
[model_providers.requesty]
name = "Requesty"
base_url = "https://router.requesty.ai/v1"
env_key = "REQUESTY_API_KEY"
wire_api = "responses"Export the key and run Codex
In WSL:
export REQUESTY_API_KEY="YOUR_REQUESTY_API_KEY"
codexIn PowerShell:
$env:REQUESTY_API_KEY = "YOUR_REQUESTY_API_KEY"
codexThe environment variable exists only in the current shell. Do not place the raw key in config.toml or commit it to a repository.
Verify the connection
Start Codex inside a small repository and ask it to explain one file. Follow with a limited edit and inspect the diff. This checks basic text generation and tool use without risking a wide change.
Kimi K3 supports images, but a text request working does not prove that the compatibility path carries image input correctly. Test image handling separately before relying on it.
Troubleshooting
A 401 error usually means REQUESTY_API_KEY is not present in the terminal that launched Codex. A model error often means the provider prefix is missing. Use moonshot/kimi-k3.
A Requesty 400 error can occur when Codex sends a reasoning summary field for a model that does not accept it. Keep model_supports_reasoning_summaries = false as a top-level setting.
Kimi K3 Codex check
- Install a current Codex CLI release
- Create and export a Requesty API key
- Use https://router.requesty.ai/v1
- Set wire_api to responses and model to moonshot/kimi-k3
- Test a small text task before checking image support
After the first small task, inspect the diff and check Requesty’s activity log. For an Anthropic-format integration, see Kimi K3 with Claude Code. For OpenCode, see Kimi K3 with OpenCode.
