operonx 0.7.0

High-performance workflow engine for AI applications

Operonx

Operonx is a workflow engine that runs anything as a workflow — from IO-bound AI tasks (LLMs, agents, RAG) to CPU-bound workloads needing native performance. Define complex pipelines as DAGs with async execution, built-in tracing, and a dual Python/Rust backend.

Why Operonx?

• DAG-based workflows — nodes and edges, inspired by Airflow operators
• Dual backend — Python for flexibility, Rust for raw speed (~8x faster on pure-compute)
• Built-in tracing — Langfuse + OpenTelemetry, plus a local viewer
• Provider agnostic — OpenAI, Azure, Gemini, Anthropic, vLLM, ONNX — swap with one line
• Type-safe state — O(1) state access with schema validation

Quick Start

pip install operonx

import asyncio
from operonx.core import Operon, GraphOp, op, START, END, PARENT

@op
def greet(name: str):
    return {"message": f"Hello, {name}!"}

async def main():
    with GraphOp(name="hello") as graph:
        step = greet(name=PARENT["name"])
        START >> step >> END

    result = await Operon(graph).run(inputs={"name": "World"})
    print(result["message"])  # Hello, World!

asyncio.run(main())

LLM Integration

pip install "operonx[standard]"

Configure resources in resources.yaml and credentials in .env, then:

import asyncio
import operonx
from operonx.core import Operon, GraphOp, START, END, PARENT
from operonx.providers import chat

async def main():
    operonx.bootstrap()  # loads ./.env + ./resources.yaml

    with GraphOp(name="chat") as graph:
        c = chat(
            resource="gpt-4o",
            template={"system": "You are a helpful assistant.", "user": "{question}"},
            question=PARENT["question"],
        )
        START >> c >> END

    result = await Operon(graph).run(inputs={"question": "What is Python?"})
    print(result["content"])

asyncio.run(main())

See Resource Setup for details on bootstrap() and resources.yaml.

Installation

Operonx is a single Python package with optional extras for each integration:

pip install operonx                    # Core engine, no providers
pip install "operonx[standard]"        # Recommended — OpenAI + Langfuse + OTEL + serve
pip install "operonx[anthropic]"       # Anthropic-only
pip install "operonx[onnx]"            # Local ONNX inference
pip install "operonx[serve]"           # FastAPI + uvicorn HTTP server
pip install "operonx[all]"             # All providers and tracers (excludes huggingface)

Extra	Contents
standard	OpenAI, Langfuse, OpenTelemetry, FastAPI/uvicorn
anthropic	Anthropic SDK
gemini	Google Vertex AI
bedrock	AWS Bedrock
onnx	ONNX Runtime + tokenizers
huggingface	transformers + torch (heavy — ~2.5 GB)
langfuse	Langfuse tracer
otel	OpenTelemetry tracer
serve	FastAPI + uvicorn HTTP server
all	Everything except huggingface
dev	pytest, ruff, mkdocs

Rust users:

cargo add operonx

Tracing

from operonx.telemetry.tracers import LangfuseTracer

engine = Operon(graph, tracer=LangfuseTracer(resource="langfuse:default"))

Backends supported: Langfuse, OpenTelemetry. Configure via resources.yaml.

Documentation

Need	Go to
Runnable examples	examples/
Architecture	docs/architecture/
User guide	docs/guide/
API reference	https://batman1m2001-cyber.github.io/Operonx/

Contributing

See CONTRIBUTING.md.

git clone https://github.com/batman1m2001-cyber/Operonx.git
cd Operonx
uv sync --all-extras
pre-commit install
uv run pytest tests/ -m "not integration"

License

Apache 2.0