nyraxis-sdk 0.6.0

Universal AI agent observability & governance — trace every LLM provider, framework, and vector DB with 17 built-in governance providers

Nyraxis Python SDK

Universal AI agent observability — trace every LLM provider, framework, and vector DB automatically.

Installation

pip install nyraxis-sdk

Quick Start (Recommended)

import nyraxis_sdk

nyraxis_sdk.init(
    api_key="nyx_your_api_key",
    base_url="https://your-nyraxis-backend.com",
    agent_name="my-agent",
)

# That's it. All LLM calls are now auto-traced.
# No code changes needed — works with any provider or framework below.

# Before process exit:
nyraxis_sdk.shutdown()

Trace Grouping

from nyraxis_sdk import workflow, task

@workflow(name="research-agent")
def run_agent(query):
    summary = summarize(query)
    return summary

@task(name="summarize")
def summarize(text):
    # LLM calls inside here are grouped under this task
    return openai.chat.completions.create(...)

Auto-Instrumented

LLM Providers (20+)

OpenAI, Anthropic, Google (Gemini/PaLM/VertexAI), Cohere, Mistral, AWS Bedrock, Azure OpenAI, Ollama, Together AI, Groq, Replicate, HuggingFace, AI21, Aleph Alpha, DeepSeek, Fireworks, Perplexity, WatsonX, SageMaker

Frameworks (10+)

LangChain, LlamaIndex, CrewAI, Haystack, AutoGen, Semantic Kernel, DSPy, Marvin, Instructor, Guidance, Letta

Vector Databases (9+)

Pinecone, Chroma, Weaviate, Qdrant, Milvus, PGVector, Redis, LanceDB, Marqo

Span Types Captured

Type	Description
llm	LLM chat/completion calls
embedding	Embedding generation
tool	Tool/function calls
retrieval	Vector DB queries
reranker	Reranking operations
agent	Agent orchestration
workflow	Workflow grouping
task	Task execution
guardrail	Guardrail checks
memory	Memory read/write
generic	Other operations

Legacy Mode (Removed)

Framework-specific handlers (framework="langchain" etc.) have been removed in v0.4.0. Auto mode covers everything they did and more. Migration:

# Before (v0.3):
handler = nyraxis_sdk.init(api_key="nyx_...", framework="langchain")
llm = ChatOpenAI(callbacks=[handler])

# After (v0.4):
nyraxis_sdk.init(api_key="nyx_...")
llm = ChatOpenAI()  # auto-traced, no callbacks needed

Direct API Client

from nyraxis_sdk import NyraxisClient

async with NyraxisClient(api_key="nyx_your_api_key") as client:
    await client.ingest_trace(
        trace_id="trace-123",
        spans=[...],
        resource_attributes={"service.name": "my-agent"},
    )

Generic OTLP Endpoint

Any OpenTelemetry-instrumented app can send traces directly:

POST /api/v1/ingest/otel/v1/traces
Header: X-API-Key: nyx_...
Content-Type: application/json
Body: Standard OTLP ExportTraceServiceRequest (JSON or Protobuf)

This means frameworks with native OTEL support (AutoGen, Semantic Kernel, DSPy, etc.) can send traces to Nyraxis without using this SDK at all.

License

MIT