Kimi CLI with Nebius Token Factory

Run Kimi CLI with Kimi K3 on Nebius Token Factory through its existing OpenAI-compatible provider. This recipe uses Kimi CLI's openai_legacy provider, which sends streaming requests to the OpenAI Chat Completions endpoint. It does not add a new provider or use the OpenAI Responses API.

The example defaults to moonshotai/Kimi-K3, Moonshot AI's reasoning and vision model with a 1M-token context window. NEBIUS_MODEL is explicit and required so the selected model is reviewed before every launch; this recipe rejects models whose context and capability declarations it has not validated.

Prerequisites

Python 3.12 or newer
uv
A Nebius Token Factory API key

The recipe was validated against Kimi CLI 1.49.0. Install the compatible 1.x line:

uv tool install --python 3.13 'kimi-cli>=1.49,<2'
kimi --version

Configure

From this directory:

cp .env.example .env

Edit .env and set both required values:

NEBIUS_API_KEY=your-token-factory-api-key
NEBIUS_MODEL=moonshotai/Kimi-K3

Load them into the current shell and validate the configuration offline:

set -a
source .env
set +a
python kimi_token_factory.py --check

The launcher validates config.template.toml, substitutes the required model into an ephemeral config, maps NEBIUS_API_KEY to the environment variable Kimi CLI's OpenAI-compatible provider reads, and deletes the generated config on exit. The API key is never written into the generated file.

Run

Start an interactive session in the current project:

python kimi_token_factory.py

Arguments not consumed by the wrapper pass through to Kimi CLI. For example:

python kimi_token_factory.py --print "Summarize this repository"

Kimi CLI remains in its normal approval mode. This recipe deliberately leaves default_yolo = false; review tool calls before approving them.

Validate the recipe

The test suite is fully offline: it checks required environment variables, parses the TOML, locks the provider to Chat Completions, verifies the canonical base URL, and ensures no secret is stored in the config.

uv venv
uv pip install -r requirements-test.txt
uv run pytest -q

Troubleshooting

`Missing required environment`

Reload .env in the shell that launches the wrapper. Both NEBIUS_API_KEY and NEBIUS_MODEL are required. This recipe requires the validated model ID moonshotai/Kimi-K3.

Authentication errors (`401` or `403`)

Create or rotate the API key in Token Factory, update .env, and reload it. The wrapper overwrites ambient OPENAI_API_KEY and OPENAI_BASE_URL values for the child process so unrelated OpenAI settings do not take precedence.

Model not found (`404`)

Model availability can change. Confirm moonshotai/Kimi-K3 in the Token Factory model catalog. If it has been replaced, update the recipe's model ID, context size, capabilities, and tests together. Do not change the provider type: Token Factory is reached through openai_legacy Chat Completions.

`Kimi CLI is not installed`

Run the installation command above and make sure the uv tools directory is on PATH (run uv tool update-shell if needed).

Search or fetch behaves differently

Kimi CLI's SearchWeb and hosted FetchURL services are exclusive to the Kimi Code platform. With Token Factory, SearchWeb is unavailable and FetchURL falls back to local fetching. This does not affect shell, file, agent, or other local tools.

How the integration works

Kimi CLI already supports arbitrary OpenAI-compatible endpoints. The essential configuration is:

[providers.token-factory]
type = "openai_legacy"
base_url = "https://api.tokenfactory.nebius.com/v1"
api_key = "overridden-by-OPENAI_API_KEY"
reasoning_key = "reasoning_content"

openai_legacy is Kimi CLI's name for the Chat Completions protocol. The separate openai_responses provider is intentionally not used by this recipe.