quartermaster-providers

Unified multi-LLM provider abstraction for Python. Write once, run against OpenAI, Anthropic, Google, Groq, xAI, or any OpenAI-compatible endpoint.

Features

• 6 Providers: OpenAI, Anthropic, Google, Groq, xAI, plus a generic OpenAI-compatible adapter
• Streaming: Async generators for token-by-token responses
• Tool Calling: Unified interface for function/tool invocation across all providers
• Structured Output: JSON-schema-constrained response generation
• Extended Thinking: Claude and o-series reasoning chains
• Vision: Image understanding on supported models
• Transcription: Audio-to-text via OpenAI Whisper
• Token Counting: Estimate tokens and cost before making requests
• Provider Registry: Register providers once, resolve by name or model pattern
• Testing Utilities: MockProvider and InMemoryHistory for unit tests
• Type-Safe: Dataclass responses with full type hints

New in v0.4.0

• CircuitBreaker -- CircuitBreaker(failure_threshold=, recovery_timeout=) wraps any provider; raises CircuitOpenError when the failure threshold is reached and recovers after the timeout.

Installation

pip install quartermaster-providers

Install with provider-specific extras:

pip install quartermaster-providers[openai]
pip install quartermaster-providers[anthropic]
pip install quartermaster-providers[openai,anthropic,google]
pip install quartermaster-providers[all]

Supported Providers

Local / Self-Hosted Providers

Register local providers with one line — the module-level helper builds a registry, normalises base_url (auto-appends /v1 if missing), honours the OLLAMA_HOST env var, and remembers the default model so callers don't have to repeat it:

from quartermaster_providers import register_local

provider_registry = register_local(
    "ollama",
    base_url="http://localhost:11434",  # or set $OLLAMA_HOST
    default_model="gemma4:26b",
)
provider = provider_registry.get("ollama")

Quick Start

Text Generation

import asyncio
from quartermaster_providers import LLMConfig
from quartermaster_providers.providers import OpenAIProvider


async def main():
    provider = OpenAIProvider(api_key="sk-...")
    config = LLMConfig(
        model="gpt-4o",
        provider="openai",
        temperature=0.7,
        max_output_tokens=1024,
    )

    response = await provider.generate_text_response(
        prompt="Explain gradient descent in two sentences.",
        config=config,
    )
    print(response.content)  # str
    print(response.stop_reason)  # "end_turn", "max_tokens", etc.


asyncio.run(main())

Tool Calling

import asyncio
from quartermaster_providers import LLMConfig, ToolDefinition
from quartermaster_providers.providers import AnthropicProvider


async def main():
    provider = AnthropicProvider(api_key="sk-ant-...")
    config = LLMConfig(model="claude-sonnet-4-20250514", provider="anthropic")

    tools = [
        ToolDefinition(
            name="get_weather",
            description="Get current weather for a location",
            input_schema={
                "type": "object",
                "properties": {
                    "location": {"type": "string", "description": "City name"},
                },
                "required": ["location"],
            },
        ),
    ]

    response = await provider.generate_tool_parameters(
        prompt="What is the weather in Tokyo?",
        tools=tools,
        config=config,
    )

    for call in response.tool_calls:
        print(f"{call.tool_name}({call.parameters})")
        # get_weather({'location': 'Tokyo'})

    print(f"Usage: {response.usage.total_tokens} tokens")


asyncio.run(main())

Tool Calling with quartermaster-tools

Tools created with @tool() integrate directly via ToolDescriptor:

from quartermaster_tools import tool


@tool()
def get_weather(city: str) -> dict:
    """Get current weather for a city.

    Args:
        city: The city name to look up.
    """
    return {"city": city, "temperature": 22}


# Convert to provider-compatible format
tool_def = get_weather.info().to_anthropic_tools()
# Or for OpenAI:
tool_def = get_weather.info().to_openai_tools()

Streaming

import asyncio
from quartermaster_providers import LLMConfig
from quartermaster_providers.providers import OpenAIProvider


async def main():
    provider = OpenAIProvider(api_key="sk-...")
    config = LLMConfig(model="gpt-4o", provider="openai", stream=True)

    async for chunk in await provider.generate_text_response(
        prompt="Write a haiku about Python.",
        config=config,
    ):
        print(chunk.content, end="", flush=True)


asyncio.run(main())

Structured Output

import asyncio
from quartermaster_providers import LLMConfig
from quartermaster_providers.providers import OpenAIProvider


async def main():
    provider = OpenAIProvider(api_key="sk-...")
    config = LLMConfig(model="gpt-4o", provider="openai")

    schema = {
        "type": "object",
        "properties": {
            "title": {"type": "string"},
            "summary": {"type": "string"},
            "topics": {"type": "array", "items": {"type": "string"}},
        },
        "required": ["title", "summary", "topics"],
    }

    response = await provider.generate_structured_response(
        prompt="Analyze the concept of reinforcement learning.",
        response_schema=schema,
        config=config,
    )

    print(response.structured_output["title"])
    print(response.structured_output["topics"])


asyncio.run(main())

API Reference

LLMConfig

Controls request behavior across all providers.

from quartermaster_providers import LLMConfig

config = LLMConfig(
    model="gpt-4o",  # Provider model identifier
    provider="openai",  # Provider name
    stream=False,  # Stream token-by-token
    temperature=0.7,  # 0.0 (deterministic) to 2.0 (creative)
    system_message=None,  # System prompt
    max_input_tokens=None,  # Input token limit
    max_output_tokens=None,  # Output token limit
    max_messages=None,  # Conversation context limit
    vision=False,  # Enable image understanding
    thinking_enabled=False,  # Extended thinking (Claude, o-series)
    thinking_budget=None,  # Max thinking tokens
    extra_body=None,  # OpenAI-compat passthrough (vLLM chat_template_kwargs, ...)
    top_p=None,  # Nucleus sampling
    top_k=None,  # Top-k sampling
    frequency_penalty=None,  # Frequency penalty (OpenAI)
    presence_penalty=None,  # Presence penalty (OpenAI)
)

On OpenAI-compatible providers (vLLM, Ollama, LM Studio, openai-compatible), thinking_enabled=False (thinking_level="off") is mapped to extra_body.chat_template_kwargs.enable_thinking=false. Qwen3.6-style chat templates enable thinking by default and ignore a top-level enable_thinking field. Official OpenAI does not get this mapping — that would attach extra_body to every ChatGPT request. Explicit extra_body.chat_template_kwargs.enable_thinking still wins.

AbstractLLMProvider Methods

Every provider implements these methods:

Cost estimation (non-abstract, returns None if pricing unavailable):

Response Types

TokenResponse -- single response or streaming chunk:

response.content  # str -- text content
response.stop_reason  # str | None -- "end_turn", "max_tokens", "tool_use"

ToolCallResponse -- tool invocation results:

response.text_content  # str -- any text alongside tool calls
response.tool_calls  # list[ToolCall] -- each has .tool_name, .tool_id, .parameters
response.stop_reason  # str | None
response.usage  # TokenUsage | None

StructuredResponse -- JSON-schema-constrained output:

response.structured_output  # dict[str, Any] -- parsed JSON
response.raw_output  # str -- raw model text
response.usage  # TokenUsage | None

NativeResponse -- complete model output:

response.text_content  # str
response.thinking  # list[ThinkingResponse] -- reasoning blocks
response.tool_calls  # list[ToolCall]
response.usage  # TokenUsage | None

TokenUsage -- token accounting:

usage.input_tokens  # int
usage.output_tokens  # int
usage.cache_creation_input_tokens  # int (Anthropic prompt caching)
usage.cache_read_input_tokens  # int
usage.total_tokens  # property: input + output

ProviderRegistry

Register providers once and resolve by name or model pattern:

from quartermaster_providers import ProviderRegistry
from quartermaster_providers.providers import OpenAIProvider, AnthropicProvider

registry = ProviderRegistry()
registry.register("openai", OpenAIProvider, api_key="sk-...")
registry.register("anthropic", AnthropicProvider, api_key="sk-ant-...")

# Get by name
provider = registry.get("openai")

# Auto-resolve from model name (gpt-* -> openai, claude-* -> anthropic, etc.)
provider = registry.get_for_model("gpt-4o")
provider = registry.get_for_model("claude-sonnet-4-20250514")

# List registered providers
registry.list_providers()  # ["anthropic", "openai"]

Model-to-provider inference patterns: gpt-*/o1-*/o3-* -> openai, claude-* -> anthropic, gemini-* -> google, llama-*/mixtral-* -> groq, grok-* -> xai.

Token Counting and Cost Estimation

from quartermaster_providers.providers import OpenAIProvider

provider = OpenAIProvider(api_key="sk-...")

tokens = provider.estimate_token_count("Hello, world!", "gpt-4o")
print(f"Estimated tokens: {tokens}")

cost = provider.estimate_cost("Hello, world!", "gpt-4o", output_tokens=100)
if cost is not None:
    print(f"Estimated cost: ${cost:.6f}")

Error Handling

All providers raise consistent exceptions from quartermaster_providers.exceptions:

from quartermaster_providers.exceptions import (
    ProviderError,  # Base exception (has .provider, .status_code)
    AuthenticationError,  # Invalid/missing API key (401)
    RateLimitError,  # Rate limited (429, has .retry_after)
    InvalidModelError,  # Model not available (404, has .model)
    InvalidRequestError,  # Malformed request (400)
    ContentFilterError,  # Blocked by safety filter (400)
    ContextLengthError,  # Input exceeds context window (400)
    ServiceUnavailableError,  # Provider temporarily down (503)
)

Testing

Use MockProvider for unit tests without real API calls:

from quartermaster_providers import LLMConfig, TokenResponse
from quartermaster_providers.testing import MockProvider

mock = MockProvider(
    responses=[
        TokenResponse(content="Paris", stop_reason="end_turn"),
        TokenResponse(content="Berlin", stop_reason="end_turn"),
    ]
)

config = LLMConfig(model="mock", provider="mock")

response = await mock.generate_text_response("Capital of France?", config)
assert response.content == "Paris"
assert mock.call_count == 1
assert mock.last_prompt == "Capital of France?"

# InMemoryHistory for conversation testing
from quartermaster_providers.testing import InMemoryHistory

history = InMemoryHistory()
history.add_message("user", "Hello")
history.add_message("assistant", "Hi there!")
assert len(history) == 2

Contributing

Contributions welcome. See CONTRIBUTING.md for guidelines.

License

Apache License 2.0. See LICENSE for details.