Providers

Provider APIs adapt STATEK requests to model-provider wire formats. Application code usually configures providers through settings; custom integrations implement LLM_API.

Symbol	Import path	Purpose	Stability
`LLM_API`	`statek.llm_api` or `statek`	Abstract provider interface	Advanced extension API
`LLM_Response`	`statek.llm_api` or `statek`	Provider response tuple	Advanced
`LLM_Stats`	`statek.llm_api` or `statek`	Usage accounting tuple	Advanced
`ModelPricing`	`statek.model_pricing`	Per-model price metadata	Advanced
`get_model_pricing`	`statek.model_pricing`	Look up or create model pricing	Advanced
`set_model_pricing`	`statek.model_pricing`	Set model pricing programmatically	Advanced
`init_model_pricing`	`statek.model_pricing`	Load pricing CSV or TXT files	Advanced
`LLM_Usage`	`statek.executors.llm_usage`	Per-job accumulated usage and cost	Advanced
`add_provider`	`statek.llm_api` or `statek`	Register a custom provider	Advanced
Built-in provider classes	`statek.llm_api` or `statek`	Typed provider wrappers	Advanced

`LLM_API`

class LLM_API(ABC):
    def preview_request(..., enable_reasoning: bool = False) -> dict
    async def process_request(..., enable_reasoning: bool = False) -> LLM_Response

Base class for provider integrations. Subclasses implement _build_request_payload(...) and _process_request(...).

Request parameters

Name	Description
`system_prompt`	Optional system prompt passed separately from chat history.
`model`	Required provider model.
`metadata`	Optional metadata. `LLM_TOOLS_SCOPE`, `TEMPERATURE`, and reasoning metadata are interpreted by request preparation paths.
`available_tools`	Tool callables available for provider tool payload selection.
`chat_history`	Iterable of `ChatHistoryItem` objects.
`chat_style`	Optional chat style that affects history formatting and tool selection.
`temperature`	Optional float from `0.0` to `1.0`.
`enable_reasoning`	Whether provider reasoning support should be requested.

Static helpers

Method	Returns	Errors
`LLM_API.get(provider_name=None, model=None, **kwargs)`	`LLM_API`	`ValueError` for missing settings or unsupported providers.
`parse_temperature(value)`	`float	None`
`require_model(model)`	`str`	`ValueError` if empty.
`parse_enable_reasoning(value)`	`bool`	No documented exception for ordinary strings.

from statek import LLM_API
 
api = LLM_API.get("OPENAI")
payload = api.preview_request(model="gpt-5-mini", system_prompt="Be concise.")

Response types

LLM_Response(text, stats, call_requests)
LLM_Stats(total_bytes_sent, total_bytes_received, cost, input_tokens, output_tokens, cached_tokens)

LLM_Response separates generated text, usage stats, and provider tool-call requests. LLM_Stats stores per-request accounting values; unavailable provider cost is represented as None. Worker execution accumulates these values into job.usage.

Metering types

ModelPricing stores optional input, cached-input, and output prices per million tokens. get_model_pricing(...), set_model_pricing(...), and init_model_pricing(...) manage pricing records used by LLM_Usage.total_cost.

LLM_Usage stores accumulated byte counts, provider token counts, cached-token counts, provider-reported cost, and calculated total_cost for a job.

See STATEK Metering for pricing file format, examples, and caveats.

Built-in provider classes

OpenAI-compatible subclasses share the default chat-completions implementation:

Provider keys	Class
`OPENAI`	`OpenAI_API`
`OPENROUTER`	`OpenRouter_API`
`GROQ`	`Groq_API`
`MISTRAL`, `MISTRALAI`, `MISTRAL_AI`	`MistralAI_API`
`DEEPSEEK`, `DEEP_SEEK`	`DeepSeek_API`
`XAI`, `X_AI`, `GROK`	`XAI_API`
`TOGETHER`, `TOGETHERAI`, `TOGETHER_AI`	`TogetherAI_API`
`FIREWORKS`, `FIREWORKSAI`, `FIREWORKS_AI`	`FireworksAI_API`
`CEREBRAS`	`Cerebras_API`
`PERPLEXITY`	`Perplexity_API`
`SAMBANOVA`, `SAMBA_NOVA`	`SambaNova_API`
`NVIDIA`, `NVIDIA_NIM`, `NIM`	`NvidiaNIM_API`
`NEBIUS`	`Nebius_API`
`COHERE`	`Cohere_API`
`MOONSHOT`, `MOONSHOTAI`, `MOONSHOT_AI`, `KIMI`	`MoonshotAI_API`
`DASHSCOPE`, `DASH_SCOPE`, `ALIBABA`, `ALIBABA_CLOUD`, `QWEN`	`DashScope_API`
`CLOUDFLARE`, `CLOUDFLARE_WORKERS_AI`, `WORKERS_AI`	`CloudflareWorkersAI_API`
`CLOUDFLARE_AI_GATEWAY`, `AI_GATEWAY`	`CloudflareAIGateway_API`
`GITHUB`, `GITHUB_MODELS`	`GitHubModels_API`
`BEDROCK`, `AMAZON_BEDROCK`, `AWS_BEDROCK`	`Bedrock_API`
`MICROSOFT_FOUNDRY`, `MS_FOUNDRY`, `AZURE_FOUNDRY`	`MicrosoftFoundry_API`
`AZURE_OPENAI`, `AZURE_OPEN_AI`	`AzureOpenAI_API`
`GEMINI_ENTERPRISE`, `GOOGLE_GEMINI_ENTERPRISE`, `GOOGLE_OPENAI`	`GeminiEnterprise_API`
`OLLAMA`	`Ollama_API`
`LMSTUDIO`, `LM_STUDIO`	`LMStudio_API`
`VLLM`	`VLLM_API`
`SGLANG`	`SGLang_API`
`LLAMA_CPP`, `LLAMACPP`, `LLAMA_CPP_PYTHON`	`LlamaCpp_API`

Specialized non-default implementations are VertexAI_API for Vertex/Gemini aliases and ClaudeAI_API for Claude/Anthropic aliases. Claude_API is exported as an alias.

`add_provider`

def add_provider(name: str, llm_api_impl: type[LLM_API] = None, **kwargs)

Registers a custom provider name. If llm_api_impl is omitted, STATEK uses the OpenAI-compatible default implementation. api_url, api_key, response_format_file, and use_prompt_caching are consumed as provider settings; other keyword arguments are passed to the provider implementation.

Raises: ValueError when the provider name is already registered or reserved.

from statek import add_provider
 
add_provider(
    "LOCAL_OPENAI_COMPAT",
    api_url="http://localhost:8000/v1/chat/completions",
    api_key="local",
)

Related APIs: Model Providers, Custom Providers, Settings.

Futures Settings