veto2 0.1.3

Real-time governance and human-in-the-loop approval for AI agents.

Veto SDK (veto2)

🛡️ Real-time governance and human-in-the-loop approval for AI agents.

Veto allows you to secure your AI agent's tool calls by intercepting them and verifying them against a central policy engine. It supports blocking, monitoring, data masking, and human-in-the-loop approvals.

Installation

pip install veto2

Quick Start

1. Initialize the SDK

Initialize Veto at the start of your application. You can provide credentials directly or use environment variables.

import veto2 as sdk

sdk.init(
    api_key="your_api_key_here",
    agent_name="SupportAgent-01"
)

Environment Variables:

• GOVERNANCE_API_KEY: Your Veto API Key.
• GOVERNANCE_AGENT_NAME: Default agent name for this instance.
• GOVERNANCE_BASE_URL: Custom backend URL (optional).

2. Protect Your Tools

Use the @sdk.guard() decorator on any function (tool) you want to monitor or control.

@sdk.guard()
def process_payment(amount: float, currency: str):
    # This code only runs if Veto policy allows it
    print(f"Processing {amount} {currency}")
    return True

How It Works

When a @sdk.guard() protected function is called:

1. Intercept: The SDK captures the function name and all arguments.
2. Verify: It sends this data to the Veto backend for policy evaluation.
3. Action: Based on the policy set in the Veto Dashboard:
   • APPROVED: The original function executes immediately.
   • BLOCKED: Raises a PermissionDeniedError with an optional feedback message.
   • PAUSE (Human-in-the-loop): The SDK pauses execution and polls the backend until a manager approves or rejects the request via the Dashboard.

Core API

sdk.init(api_key, agent_name, base_url=None)

Initializes the global governance client.

• api_key: (Required) Found in your Veto Dashboard settings.
• agent_name: (Optional) Descriptors for your agent. Defaults to "DefaultAgent".
• base_url: (Optional) The API endpoint of your Veto backend.

@sdk.guard(name=None)

The primary decorator for tool protection.

• name: (Optional) Override the tool name sent to the dashboard. Defaults to the function's name.

sdk.ping()

Utility function to verify the connection to the Veto backend and validate the API key.

Policy Modes (Configured via Dashboard)

Veto is designed to be "Code-First, Policy-Second". You wrap your functions once, and manage behavior from the UI:

Mode	Behavior
BLOCK	Denies the request immediately.
SHADOW	Allows the request but logs it as a warning for monitoring.
HUMAN_APPROVAL	Pauses execution until a person clicks "Approve" in the Dashboard.
MASKING	Redacts sensitive parameters from logs while allowing the original call.
FEEDBACK	Blocks the call and returns a specific hint/instruction to the LLM.
RULES	Dynamically chooses a mode based on argument values (e.g., amount > 1000).

Error Handling

Veto uses specific exceptions to help you manage governance violations:

from veto2 import PermissionDeniedError

try:
    sensitive_operation()
except PermissionDeniedError as e:
    # Log the violation or inform the user/LLM
    print(f"Action blocked by governance: {e}")

Plan-Based Timeouts (Human Approval)

When an action is paused for human review, the SDK will poll for a decision:

• Basic Plan: Polls for up to 3 hours.
• Pro Plan: Polls for up to 7 days.

---

Built for secure AI agent deployment. Get your API key