astrai-sdk 0.2.0

AI inference in your terminal. PII-stripped, compliance-logged, cost-optimized.

Astrai — Compliant LLM Gateway

PII stripping, audit logs, multi-model routing. Built for regulated industries.

pip install astrai

Quick Start

# Configure your API key
astrai config --api-key YOUR_KEY

# Chat with the best model (auto-routed)
astrai chat "What is risk parity?"

# Use a specific model
astrai chat -m claude-opus-4.5 "Explain Black-Scholes"

# Export audit log for compliance
astrai audit --days 30 -o audit.json

Commands

Command	Description
astrai chat	Send a chat message (streaming)
astrai models	List available models and pricing
astrai health	Model health dashboard
astrai usage	Usage statistics and spend
astrai audit	Export compliance audit log
astrai preferences	View learned routing preferences
astrai config	Configure API key, model, strategy

Features

• Auto-routing: Automatically picks the best model for cost/quality/latency
• PII stripping: Sensitive data is stripped before reaching external models
• Audit logs: Every request is logged with model, cost, latency, routing reason
• BYOK: Bring your own API keys per provider
• Fallback chains: Claude -> GPT -> Llama automatic failover
• EU routing: Route to EU-based providers for GDPR compliance

Models

Tier	Models	Use Case
Ultra	claude-opus-4.5, gpt-5.2, gemini-3-pro	Complex analysis
Target	llama-4-maverick, deepseek-v3.2, llama-3.3-70b	General use
Draft	llama-4-scout, gemini-3-flash	Fast/cheap tasks
Reasoning	o3, deepseek-r1	Multi-step reasoning

Environment Variables

Variable	Description
ASTRAI_API_KEY	Your Astrai API key
ASTRAI_API_URL	Custom API endpoint
ASTRAI_MODEL	Default model override

For Compliance Officers

Export a complete audit trail of all AI model usage:

# JSON export for the last 30 days
astrai audit --days 30 --format json -o compliance_report.json

# Weekly CSV export
astrai audit --days 7 --format csv -o weekly_audit.csv

Each audit record includes:

• Timestamp, model used, provider
• Routing reason (cost, quality, preference)
• Token count, cost, latency
• Task type classification