airg-client 0.3.1

Python client for the AI Runtime Governor – runtime AI-agent governance service

airg-client

Python client for the AI Runtime Governor – a runtime governance layer for AI agents.

The Governor evaluates every tool invocation against configurable policies and returns an allow / block / review decision before the tool executes.

Installation

pip install airg-client

Quick start

The client reads GOVERNOR_URL and GOVERNOR_API_KEY from environment variables automatically — just like the OpenAI SDK.

from governor_client import evaluate_action, GovernorBlockedError

try:
    decision = evaluate_action(
        tool="shell_exec",
        args={"command": "ls -la /tmp"},
        context={"session_id": "abc-123"},
    )
    print(decision["decision"])      # "allow" | "review"
    print(decision["risk_score"])    # 0.0 – 1.0
    print(decision["explanation"])
except GovernorBlockedError as e:
    print(f"Blocked: {e}")

Configuration

Environment variable	Default	Description
GOVERNOR_URL	http://localhost:8000	Base URL of the Governor service
GOVERNOR_API_KEY	(empty)	API key (airg_…) sent as X-API-Key header

You can also override them in code:

import governor_client

governor_client.GOVERNOR_URL = "..."
governor_client.GOVERNOR_API_KEY = "airg_..."

API

evaluate_action(tool, args, context=None) → dict

Send a tool call to the Governor for evaluation.

Returns the full decision dict:

{
    "decision": "allow",        # "allow" | "block" | "review"
    "risk_score": 0.15,
    "explanation": "Low-risk read operation",
    "policy_ids": ["shell-guard"],
    "modified_args": None,
}

Raises GovernorBlockedError if the decision is "block".

governed_call(tool, args, context=None) → dict

Convenience wrapper around evaluate_action. Identical behaviour — callers should inspect decision for "review" and handle accordingly.

GovernorBlockedError

Exception raised when the Governor blocks a tool invocation. Subclass of RuntimeError.

Requirements

• Python ≥ 3.9
• httpx ≥ 0.24.0

License

MIT — see LICENSE.