Moonshot and Kimi Code

Rho supports Moonshot's OpenAI-compatible Chat Completions API through two explicit provider selections. Neither selection uses the Anthropic Messages API.

At a glance

Moonshot

Setting	Value
Provider	moonshot
Auth	moonshot-api-key
Environment override	MOONSHOT_API_KEY
API base	https://api.moonshot.ai/v1
Model list	Refreshable after authentication

Kimi Code

Setting	Value
Provider	kimi-code
Auth	kimi-oauth
Environment override	KIMI_ACCESS_TOKEN
API base	https://api.kimi.com/coding/v1
Model list	Refreshable after authentication

Both providers send turns to /chat/completions and fetch models from /models. Their model lists can be updated through Refresh model lists in /config after authentication. Rho preserves account-specific context limits returned by the model-list API. For Kimi Code, account-specific reasoning choices replace general models.dev reasoning metadata when available.

Although Moonshot and Kimi Code share an OpenAI-compatible transport, they use separate reasoning contracts:

• Moonshot Kimi K3 requests use the top-level reasoning_effort field.
• Kimi Code K3 requests use thinking: { "type": "enabled", "effort": "<level>" }. Turning reasoning off sends thinking: { "type": "disabled" } without an effort.

Rho does not send Anthropic-style adaptive thinking or output_config.effort to either provider.

Moonshot API key

In the interactive TUI, run:

/login moonshot
/model moonshot/<model-id>

Rho stores the key in the configured credential store. For CI and development, MOONSHOT_API_KEY overrides the stored key. Remove the stored key with /logout moonshot.

For automation, select the matching provider and auth mode:

rho --provider moonshot --auth moonshot-api-key --model <model-id> run "hello"

Kimi Code OAuth

Kimi Code uses the RFC 8628 device authorization flow implemented by MoonshotAI/kimi-code. Run either of these commands:

/login kimi-code

rho login kimi-code

Visit the displayed Kimi authorization URL and enter the displayed code if requested. Rho stores the access and refresh tokens in the configured credential store, refreshes expiring tokens, and persists replacements. KIMI_ACCESS_TOKEN is a non-persistent CI/development override and cannot be refreshed.

Select a model after login:

/model kimi-code/<model-id>

For Kimi K3, Rho uses the authenticated context_length returned for the current account. If that field is unavailable, Rho uses a conservative 256K effective window while retaining the model's advertised 1M maximum. K3 reasoning choices come from the authenticated model list when available. Enabled requests carry the selected effort in Kimi Code's native thinking object. Turning reasoning off sends the native disabled mode, which Kimi documents as routing the request to K2.6.

For automation after a prior login or with KIMI_ACCESS_TOKEN:

rho --provider kimi-code --auth kimi-oauth --model <model-id> run "hello"

Use /limits in the TUI to inspect the weekly and rolling usage windows reported by Kimi Code. Remove stored OAuth tokens with /logout kimi-code. An active KIMI_ACCESS_TOKEN override remains available after logout.