sdkrouter 0.1.32

Unified SDK for AI services with OpenAI compatibility

SDKRouter

Unified Python SDK for AI services. Access 300+ LLM models, vision, audio, image generation, search, translation, and more through a single interface.

Installation

pip install sdkrouter

Quick Start

from sdkrouter import SDKRouter, Model

client = SDKRouter(api_key="your-api-key")

response = client.chat.completions.create(
    model=Model.cheap(),
    messages=[{"role": "user", "content": "Hello!"}]
)
print(response.choices[0].message.content)

Features

Feature	Description	Docs
Chat	OpenAI-compatible completions, streaming	@docs/01-chat.md
Structured Output	Pydantic models, JSON extraction	@docs/02-structured-output.md
Audio	TTS, STT, Deepgram streaming	@docs/03-audio.md
Vision	Image analysis, OCR	@docs/04-vision.md
Image Gen	AI image generation	@docs/05-image-gen.md
Search	Web search with modes	@docs/06-search.md
CDN	File storage	@docs/07-cdn.md
Translator	JSON/text translation	@docs/08-translator.md
Payments	Crypto payments	@docs/09-payments.md
Proxies	Proxy management	@docs/10-proxies.md
Embeddings	Text embeddings	@docs/11-embeddings.md
Other	Shortlinks, cleaner, models API	@docs/12-other.md

Model Routing

Smart model selection with IDE autocomplete:

from sdkrouter import Model

Model.cheap()                    # Lowest cost
Model.smart()                    # Highest quality
Model.balanced()                 # Best value
Model.fast()                     # Fastest

# With capabilities
Model.cheap(vision=True)         # + vision
Model.smart(tools=True)          # + function calling
Model.balanced(json=True)        # + JSON mode

# Categories
Model.smart(code=True)           # Coding
Model.cheap(reasoning=True)      # Problem solving

Async Support

from sdkrouter import AsyncSDKRouter, Model
import asyncio

async def main():
    client = AsyncSDKRouter(api_key="your-api-key")

    response = await client.chat.completions.create(
        model=Model.cheap(),
        messages=[{"role": "user", "content": "Hello!"}]
    )

    # Parallel requests
    results = await asyncio.gather(
        client.vision.analyze(image_url="..."),
        client.audio.speech(input="Hello!"),
    )

asyncio.run(main())

Audio Example

from sdkrouter import SDKRouter, AudioModel

client = SDKRouter()

# Text-to-Speech
response = client.audio.speech(
    input="Hello!",
    model=AudioModel.cheap(),
    voice="nova",
)
Path("output.mp3").write_bytes(response.audio_bytes)

# Speech-to-Text
result = client.audio.transcribe(file=audio_bytes)
print(result.text)

Deepgram Streaming

from sdkrouter import AsyncSDKRouter
from sdkrouter.tools.audio.stt import DeepgramConfig

sdk = AsyncSDKRouter()

config = DeepgramConfig(
    model="nova-3",
    endpointing=300,   # VAD: silence threshold (ms)
    vad_events=True,   # Enable VAD events
)

async with sdk.audio.stt.stream_deepgram(config) as session:
    await session.send(audio_chunk)
    async for segment in session.transcripts():
        print(segment.text)

Provider Selection

By default the server auto-detects the provider from the model name. You can override this explicitly:

# Client-level — all requests go through this provider
client = SDKRouter(api_key="your-key", provider="openrouter")

# Per-request override via extra_body
response = client.chat.completions.create(
    model="qwen3-max",
    messages=[{"role": "user", "content": "Hi"}],
    extra_body={"provider": "alibaba"},  # overrides client-level
)

# Works with async too
client = AsyncSDKRouter(api_key="your-key", provider="alibaba")
result = await client.parse(
    model="qwen3-max",
    messages=[...],
    response_format=MyModel,
)

Available providers: openrouter, openai, anthropic (more coming).

Direct Provider Routing

By default (use_self_hosted=True) all LLM and embedding requests go through llm.sdkrouter.com. To route directly to a provider using your own API key:

# OpenAI directly
client = SDKRouter(
    api_key="sk-...",
    llm_url="https://api.openai.com/v1",
    use_self_hosted=False,
)

# OpenRouter directly
client = SDKRouter(
    api_key="sk-or-...",
    llm_url="https://openrouter.ai/api/v1",
    use_self_hosted=False,
)

In both cases api_url (CDN, vision, search, shortlinks) stays on api.sdkrouter.com — only LLM/embeddings traffic is redirected to llm_url.

Embeddings

# Via sdkrouter proxy (default, single key)
client = SDKRouter(api_key="your-sdkrouter-key")
result = client.embeddings.create(
    ["Hello world", "Semantic search"],
    model="openai/text-embedding-3-small",
)
vectors = [item.embedding for item in result.data]

# Directly via OpenAI key
client = SDKRouter(
    api_key="sk-...",
    llm_url="https://api.openai.com/v1",
    use_self_hosted=False,
)
result = client.embeddings.create("Hello world", model="text-embedding-3-small")

# Directly via OpenRouter key
client = SDKRouter(
    api_key="sk-or-...",
    llm_url="https://openrouter.ai/api/v1",
    use_self_hosted=False,
)
result = client.embeddings.create("Hello world", model="openai/text-embedding-3-small")

Model	Dimensions
openai/text-embedding-3-small	1536 (default)
openai/text-embedding-3-large	3072
openai/text-embedding-ada-002	1536 (legacy)

Configuration

# Environment variables (auto-loaded)
# SDKROUTER_API_KEY
# SDKROUTER_LLM_URL
# SDKROUTER_API_URL
# SDKROUTER_AUDIO_URL

client = SDKRouter(
    api_key="your-key",
    timeout=60.0,
    max_retries=3,
)

Prompt Caching & Metrics

For Anthropic Claude models, SDKRouter automatically applies cache_control breakpoints. Cache metrics are returned in response.usage.prompt_tokens_details:

response = client.chat.completions.create(
    model="anthropic/claude-haiku-4-5",
    messages=[...],  # long conversation
)

details = response.usage.prompt_tokens_details
if details:
    print("Cache read tokens: ", details.cached_tokens)      # billed at 10%
    print("Cache write tokens:", details.cache_write_tokens) # billed at 125%

No client-side changes needed — caching is transparent and automatic.

Supported Providers

• OpenAI: GPT-4.5, GPT-4o, o3, o1
• Anthropic: Claude Opus 4.6, Claude Sonnet 4.6, Claude Haiku 4.5
• Google: Gemini 2.5 Pro, Gemini 2.0 Flash
• Alibaba DashScope: Qwen3-Max, Qwen3.5-Plus, Qwen-Plus, QwQ-32B, Qwen3-VL
• Meta: Llama 4, Llama 3.3
• Mistral: Mistral Large, Codestral
• DeepSeek: DeepSeek V3, R1
• And 300+ more via OpenRouter

License

MIT