Skip to content

chimera.providers

chimera.providers is the LLM-backend layer. Every provider implements the Provider ABC; create_provider() auto-detects the right one from a model name or explicit provider_type.

Top-level exports

from chimera.providers import (
    Provider,
    Response,
    StreamEvent,
    create_provider,
    register_provider,
    get_provider_factory,
    list_providers,
    ThinkingLevel,
    budget_for_level,
    ProxyProvider,
)
from chimera.providers.catalog import ModelConfig, ProviderCatalog

Symbol	Purpose
`Provider`	ABC. Implement `complete(messages, tools=...)` (sync) and/or `stream(...)`. Accepts a `thinking` parameter.
`Response`	Dataclass with `content`, `tool_calls`, `usage`, `cost`, `stop_reason`.
`StreamEvent`	Streaming delta type yielded by `Provider.stream`.
`create_provider(provider_type=None, model=None, **kwargs)`	Factory with model-name autodetection (see below).
`register_provider(name, factory)`	Register a custom provider type at runtime.
`get_provider_factory(name)` / `list_providers()`	Inspect the registry.
`ThinkingLevel`	Enum: `OFF` / `MINIMAL` / `LOW` / `MEDIUM` / `HIGH` / `MAX`.
`budget_for_level(level)`	Map a `ThinkingLevel` to a token budget.
`ProxyProvider`	Wrap another provider to intercept/log/transform requests.
`ModelConfig` / `ProviderCatalog`	Catalog API for slash-namespaced models (`bedrock/claude-sonnet-4`, `azure/gpt-4o`, `groq/llama-3.3-70b`, …).

Built-in provider modules

Module	Class	Routes by name prefix
`chimera.providers.anthropic`	`AnthropicProvider`	`claude-*`
`chimera.providers.openai_provider`	`OpenAIProvider`	`gpt-`, `o1-`, `o3-*`
`chimera.providers.google`	`GoogleProvider`	`gemini-*`
`chimera.providers.ollama`	`OllamaProvider`	`llama`, `mistral`, `qwen`, `phi`
`chimera.providers.modal`	`ModalProvider`	(explicit)
`chimera.providers.compatible`	`CompatibleProvider`	OpenAI-compatible endpoints (vLLM, LiteLLM, Together, Groq, Fireworks). Requires `base_url`.
`chimera.providers.proxy`	`ProxyProvider`	Wraps any other provider.

`create_provider()` resolution order

Explicit provider_type (e.g. "anthropic", "compatible").
Known model-name prefix (the table above).
ProviderCatalog.default() lookup (slash-namespaced ids).
Environment-variable fallback (ANTHROPIC_BASE_URL / ANTHROPIC_AUTH_TOKEN → Anthropic; OPENAI_API_KEY → OpenAI).

For the full list of catalog ids and their env-var contracts (DeepSeek, GLM, Kimi, Qwen3, GPT-OSS, Mistral Codestral, Gemma 3, Bedrock, Azure, Groq), see Use with Third-Party Providers.

Cost tracking

from chimera.providers.cost import register_model_cost
from chimera.providers.cost_tracker import CostTracker

register_model_cost(model, input_per_mtok, output_per_mtok) plugs a custom price into the global table; CostTracker records granular per-step token counts (cache hits, reasoning tokens, regular input / output) and rolls them up into a session total.

See also

Use with Third-Party Providers for end-to-end env-var setups.
modules/provider-registry for the registry’s behaviour with plugin-supplied providers.
chimera.providers.thinking for the extended-thinking budget table.