How to use Kimi K3 in Codex CLI

Codex CLI is a coding agent, not a model lock-in. When its provider uses an OpenAI-compatible Responses endpoint, the agent workflow can run against another available model. AnyModel uses that compatibility layer to expose Kimi K3 in Codex with the model id kmc/k3.

Kimi K3 is Moonshot's flagship route for complex coding and knowledge work. On AnyModel it uses connected Kimi Code capacity with a 256K context limit. You authenticate and pay with an AnyModel key and balance; you do not receive a direct Kimi Code account.

Fast setup: install with Kimi K3 selected

Create an AnyModel key, then run:

bash <(curl -fsSL "https://anymodel.org/i?tool=codex&model=kmc/k3") <YOUR_API_KEY>

The model query parameter is important because the served id contains the kmc/ provider prefix. The installer sanitizes the value and writes it to the local Codex provider configuration.

After installation, open a repository with a recoverable git state and run a small task:

codex --model kmc/k3 "Inspect the failing tests, explain the likely cause, and do not edit yet."

Starting with analysis-only work confirms the route before the agent changes files.

What the installer configures

Codex agent features use the Responses wire format. A manual provider block should contain the same essential values:

model = "kmc/k3"
model_provider = "anymodel"

[model_providers.anymodel]
name = "AnyModel"
base_url = "https://anymodel.org/v1"
wire_api = "responses"
requires_openai_auth = true

The API key is stored through Codex authentication rather than embedded in a repository config. Never commit a real key or paste it into a public support message.

If you have already configured AnyModel for GPT-5.6, you do not need a second endpoint or key. Override the model for a single session:

codex --model kmc/k3 "Implement the parser change and run the focused test suite."

When to try Kimi K3 in Codex

Kimi K3 is worth testing on tasks that need more than autocomplete:

• refactors across several modules;
• debugging that combines logs, code and tests;
• long plans with repeated shell or repository tools;
• game systems and state-heavy application logic;
• code review where the model must connect distant dependencies.

Its 256K context is large, but smaller than the 1M windows of GPT-5.6 Sol and Terra. The correct comparison is not the maximum number alone. Give each model the relevant files, same task definition and same tests, then compare the patch and corrections.

The Kimi K3 vs GPT-5.6 Sol comparison shows catalog facts side by side. For pricing and route details, read the Kimi K3 API guide.

A two-model Codex workflow

One endpoint makes deliberate routing easy:

1. Use GPT-5.6 Terra as the normal baseline.
2. Run Kimi K3 on tasks where you want an independent implementation or review.
3. Escalate the hardest failures to GPT-5.6 Sol.
4. Use a cheaper worker for mechanical follow-up changes.

For a fair second opinion, reset to the same commit or use separate worktrees. If the second model sees the first model's patch, you are measuring review behavior, not independent problem solving.

Troubleshooting

Model not found. Use kmc/k3, including the slash. kimi-k3 is the public page slug, not the served API id. Confirm the current id on the model page.

401 authentication error. Make sure the provider points to AnyModel and Codex is reading the AnyModel key. A Kimi account key, OpenAI key and AnyModel key are different credentials.

Requests fail after working earlier. Kimi K3 depends on subscription-backed upstream capacity. Retry transient failures with backoff and keep a fallback model for work that must finish on schedule.

Codex behaves like chat instead of an agent. Check wire_api = "responses". A generic chat-completions provider may return text without matching the Responses tool loop Codex expects.

Large tasks hit context or become unstable. Limit the file set, ask Codex to search before reading entire directories and split implementation from review. The 256K limit applies to a request context; a session can make many requests.

Regional access or payment is unclear. The Kimi K3 in Russia page explains the independent API route, checkout-dependent payment methods and network caveats.

Repository and privacy safeguards

Before enabling edits, commit current work or create a separate worktree. State which paths the agent may touch, require tests and review the diff. Secrets, production dumps and private keys should never enter the prompt.

Ghost Mode provides zero retention on the AnyModel side. It does not prevent the upstream model from processing the request, so match the data you send to your own security policy.

Kimi K3 in Codex takes one installer command, but evaluate it like an engineering dependency: verify the id, test a representative task, measure the complete session and keep a fallback. Create an AnyModel key to compare Kimi K3 and GPT-5.6 from the same Codex setup.