How to Configure Different LLM Providers Other Than the Default in Kimi CLI config.toml

You configure alternative LLM providers by defining [providers.<name>] tables in your ~/.kimi/config.toml file, specifying the type and api_key, then either setting default_provider globally or passing the --provider flag for individual commands.

The Kimi CLI (MoonshotAI/kimi-cli) uses a TOML-based configuration system to manage connections to various LLM backends. To configure different LLM providers other than the default in config.toml, you edit the provider-specific tables that the Config Pydantic model parses from src/kimi_cli/config.py. These definitions are validated by src/kimi_cli/web/api/config.py and instantiated into runtime objects by src/kimi_cli/llm.py.

Provider Configuration Schema

Each provider is defined as a sub-table under the top-level providers key. The CLI uses these entries to build provider instances through the kosong chat-provider abstraction layer.

Required fields for every provider entry:

• type – The provider implementation name (e.g., openai, anthropic, kimi).
• api_key – The authentication secret for the service.

Optional fields include:

• model – Default model name to use when this provider is selected.
• base_url – Custom endpoint URL for self-hosted or proxy deployments.
• custom_headers – Additional HTTP headers required by certain services.
• oauth – OAuth configuration block for managed authentication flows.

Configuring Cloud Provider Examples

OpenAI

Add the following table to ~/.kimi/config.toml:

[providers.openai]
type = "openai"
api_key = "sk-your-openai-key-here"
model = "gpt-4o"

Anthropic

[providers.anthropic]
type = "anthropic"
api_key = "sk-ant-your-anthropic-key"
model = "claude-3-5-sonnet-20240620"

Moonshot AI (Kimi)

[providers.kimi]
type = "kimi"
api_key = "YOUR_KIMI_API_KEY"
model = "kimi-latest"

For OAuth-managed keys (where tokens are refreshed via the KIMI platform), use the oauth block instead of a static api_key:

[providers.kimi_managed]
type = "kimi"
oauth = { client_id = "abc123", redirect_uri = "http://localhost:8000/callback", scopes = ["chat"] }

Self-Hosted and Custom Endpoints

To point the CLI at a self-hosted OpenAI-compatible server, specify the base_url field:

[providers.local_llm]
type = "openai"
api_key = "dummy-key-for-local"
base_url = "http://localhost:8000/v1"
model = "llama3.1-70b"

When src/kimi_cli/llm.py constructs the LLMRuntime, it passes this base_url to the underlying provider factory, overriding the default service endpoint.

Switching Between Providers

Global Default

Set the default_provider key at the root of config.toml to specify which provider loads automatically:

default_provider = "anthropic"

This value is read by the Config class in src/kimi_cli/config.py and used as the fallback when no other selection is made.

Per-Command Override

As implemented in src/kimi_cli/cli/__init__.py, you can override the default for a single execution using the --provider flag:

kimi --provider openai chat "Explain quantum mechanics"
kimi --provider local_llm chat "Summarize this document"

This flag takes precedence over the default_provider setting in the TOML file.

Summary

• Provider definitions reside in [providers.<name>] tables inside ~/.kimi/config.toml.
• Required fields are type and api_key; optional fields include model, base_url, custom_headers, and oauth.
• Global defaults are controlled by the top-level default_provider key.
• Runtime switching is supported via the --provider CLI argument managed in src/kimi_cli/cli/__init__.py.
• Implementation chain: src/kimi_cli/config.py loads the file, src/kimi_cli/web/api/config.py validates the schema, and src/kimi_cli/llm.py instantiates the runtime.

Frequently Asked Questions

Where does Kimi CLI look for the config.toml file?

By default, the CLI reads ~/.kimi/config.toml. You can specify an alternative path using the --config command-line argument, which the parser handles in src/kimi_cli/cli/__init__.py.

Can I store API keys for multiple providers at once?

Yes. Define separate [providers.<name>] sections for each service. The CLI only uses the provider matching the active selection (either default_provider or the value passed to --provider), leaving other configurations dormant but available for instant switching.

How do I configure a provider that requires custom HTTP headers?

Add a custom_headers table under your provider entry containing key-value pairs for the required headers. The ProviderConfig model in src/kimi_cli/web/api/config.py parses these and injects them into the HTTP client session used by the kosong provider abstraction.

Is it possible to use OAuth instead of a static API key?

Yes. Replace the api_key field with an oauth block containing client_id, redirect_uri, and scopes. The CLI triggers the OAuth flow on first use and caches refreshed tokens in the session store, as handled by the authentication logic in src/kimi_cli/web/api/config.py.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}