$ provara --approve "Get-Process python"

Secure AI Agent
Command Orchestration

Human-in-the-loop approval for every command your AI agents execute. Policy engine. Audit trail. Zero trust by default.

v0.1.0 · MIT License · Python 3.12+

How It Works

Every command flows through a secure approval pipeline before touching your system.

┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│  AI Agent   │────▶│  /plan API  │────▶│   Pending   │
│ (AutoGen,   │     │  (FastAPI)  │     │   Queue     │
│  Claude,    │     └─────────────┘     └──────┬──────┘
│  Custom)    │                                
└─────────────┘                                
┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│   Audit     │◀────│   Policy    │◀────│  Human UI   │
│    Log      │     │   Engine    │     │  (Approve)  │
└─────────────┘     └─────────────┘     └─────────────┘

Built for Security-First AI Operations

Every layer designed to keep humans in control of AI agent actions.

Human-in-the-Loop

Every command queued for human review before execution. No autonomous access to your system — ever.

Policy Engine

26 deny rules block dangerous operations. Allow patterns match command intent. Deny-first, category-based filtering.

Full Audit Trail

Every command logged with timestamp, result, duration, and exit code. JSON-line audit log for compliance and forensics.

Agent SDK

Python SDK and PowerShell modules. Native AutoGen integration. Build custom agents with one function call.

Real-Time Events

Server-Sent Events for live run status. Watch commands queue, execute, and complete in real time.

File System Jail

Working directory restrictions and path validation. File API jailed to project root. No escape to system directories.

One Function. Full Safety.

Integrate Provara into any Python agent with a single import.

agent.py
from provara import plan

# Queue a command for human approval
result = plan(
    command="Get-Process python",
    note="Checking running Python processes",
    timeout_s=30
)

print(f"Queued: {result['pending_id']}")
# → Queued: 20250213_143022_a1b2c3d4
# Command waits for human approval before executing

Ready to secure your AI agents?

Open source. MIT licensed. Deploy in minutes.