Providers and Models¶

Kerkov works with the model you prefer. Seven providers are supported, and you switch between them — and between models within a provider — at runtime without changing anything else about the workflow.

Provider	Aliases	Authentication
`anthropic`	`claude`	API key
`gemini`	`google`, `google-gemini`	API key
`nvidia`	`nim`, `nvidia-nim`	API key (NVIDIA NIM)
`openrouter`	`open-router`	API key
`deepseek`	—	API key
`ollama`	—	Local server URL; no key by default
`codex`	`openai-codex`	Browser-based OAuth sign-in

Choosing a provider¶

Set the startup provider with LLM_PROVIDER (default gemini), or switch in a session:

kerkov › /provider
kerkov › /provider claude

With no argument, /provider opens a picker and highlights the active one. Selecting codex starts browser sign-in when it's needed; the resulting tokens are stored under CODEX_STATE_PATH and refreshed automatically.

Choosing a model¶

Each provider has a default model (see the configuration reference). Change it with /model, or set a DEFAULT_<PROVIDER>_MODEL variable:

kerkov › /model
kerkov › /model claude-sonnet-4-5-20250929

The picker lists what the current provider offers; switching provider resets the model to that provider's default unless you pick another.

Credentials¶

API keys are entered at a masked prompt — with /auth, or automatically the first time you select a provider that needs one — so they never enter the conversation or logs. Kerkov validates a key before saving it to your OS keyring (or a user-only file on headless systems).

Environment variables remain available for CI and non-interactive use: set ANTHROPIC_API_KEY, GOOGLE_API_KEY, NVIDIA_API_KEY, OPENROUTER_API_KEY, DEEPSEEK_API_KEY, and the rest as listed in the configuration reference.

Remove a stored credential with /logout.

In the web console, Manage providers offers API-key entry, connection testing, and Ollama endpoint settings. Use Save as default in the model picker to persist a startup provider/model. Codex sign-in still starts through the CLI.

Provider notes¶

Ollama talks to a local server at OLLAMA_BASE (http://localhost:11434 by default) and needs no key. Point it at a remote Ollama by changing that URL.
NVIDIA NIM, OpenRouter, and DeepSeek each have a *_BASE_URL override for self-hosted or proxied endpoints. OpenRouter also accepts OPENROUTER_HTTP_REFERER and OPENROUTER_APP_TITLE for app attribution.
Codex uses OAuth rather than a key; static client configuration lives in config/codex.json (copy from config/codex.example.json).

Context budget¶

Conversation length is bounded by MAX_CONTEXT_TOKENS (default 50000) and per-response output by MAX_RESPONSE_TOKENS; compaction runs at COMPACTION_THRESHOLD_PCT (65%) of the budget. Raise these for longer conversations, within the selected model's limits and your cost budget. What survives compaction is covered under Memory.