Aurex is an in-process AI runtime layer designed for production LLM applications and autonomous agent frameworks. Embedded directly within your application host process, Aurex intercepts LLM provider calls with under 2 milliseconds of overhead to handle model routing, context compression, behavioral loop detection, and automatic provider failover.

Key Features

• Sub-2ms In-Process Execution: All decisions, loop guards, and routing rules execute locally in memory with zero network hop latency.
• Universal Auto-Patching: Single-line integration for OpenAI, Anthropic, Google Gemini, LiteLLM, Ollama, and Hugging Face.
• Zero Prompt Data Egress: Operates local-first using SHA-256 signatures. Sensitive prompt payloads and LLM responses never leave your application host.
• Adaptive Model Routing: Dynamically routes low-complexity calls to higher-efficiency model tiers (for example, GPT-4o to GPT-4o-mini).
• Behavioral Loop Mitigation: Hybrid structural and semantic detection (F1 score: 0.72) stops infinite agent loops and recursive tool cascades.

Installation

pip install aurex-sdk

For Node.js environments:

npm install aurex-sdk

Quick Start Guide

Option A: Zero-Code Auto-Hook (Recommended)

Attach Aurex runtime hooks at application launch with no modifications to existing codebase logic.

# Register global site-packages import hooks
aurex install-hook

# Configure environment variables and start your application
export AUREX_ENABLED=1
export AUREX_API_KEY=aux_dev_yourkey
python your_app.py

Option B: Programmatic SDK Setup

Initialize the Aurex auditor singleton and auto-patch your LLM clients directly.

import openai
from aurex_sdk import AurexAuditor, patch_all

# Initialize the auditor and patch installed providers
auditor = AurexAuditor()
patch_all()

# Use standard provider clients as usual
client = openai.OpenAI()
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Explain quantum computing in simple terms."}]
)

# Access runtime health and efficiency metrics
report = auditor.get_score()
print(f"Efficiency Score: {report.value}/100")

Option C: Runtime Guards and Circuit Breakers

Enforce spending limits and pre-call policy routing before sending API calls to providers.

from aurex_sdk import AurexAuditor

auditor = AurexAuditor()
auditor.configure_circuit_breaker({
    "max_cost_per_minute": 5.00,
    "max_cost_per_session": 50.00,
    "on_budget_exceeded": "throw"  # Options: "throw", "warn", "webhook"
})

# Pre-call evaluation
decision = auditor.pre_call_check(
    model="gpt-4o",
    prompt="Generate system report.",
    workflow_id="session_4921"
)

if decision.action == "block":
    raise Exception(f"Call blocked: {decision.reason}")
elif decision.action == "downgrade":
    model = decision.suggested_model  # Route to suggested efficient model

Offline CI/CD Quality Gates

Enforce cost and efficiency thresholds during continuous integration builds:

# Scan local ledger for optimization opportunities and fail CI build on policy breach
aurex audit --ledger .aurex/ledger.jsonl --fail-under 85 --mode block

Self-Hosting and Dashboard Setup (Optional)

For local development and self-hosted deployments, telemetry flushes asynchronously to an optional Aurex backend and analytics dashboard.

1. Backend Server: FastAPI backend service at https://heyaurex.com/api/v1.
2. Analytics Dashboard: Next.js UI running at https://heyaurex.com.

To run the local control plane during development, clone the repository and refer to the full deployment documentation at https://heyaurex.com/docs.

License and Documentation

• Interactive Documentation: https://heyaurex.com/docs
• GitHub Repository: https://github.com/bhav09/aurex-sdk