codesec-agentguard (Python SDK)

Official Python SDK for the AgentGuard AI Agent Security Gateway.

AgentGuard intercepts, evaluates, and verifies autonomous AI agent actions against deterministic spending limits, currency controls, capability whitelists, and adversarial prompt injection heuristic scans before downstream payment APIs (such as Razorpay or PayPal) are executed.

Installation

pip install codesec-agentguard

Or install locally in editable mode:

pip install -e packages/python-sdk

Quickstart

import os
from agentguard import AgentGuard

# Initialize client
guard = AgentGuard(
    base_url=os.environ.get("AGENTGUARD_URL", "http://localhost:3001"),
    api_key=os.environ["AGENTGUARD_API_KEY"],
    timeout=5.0,
)

# Pre-flight evaluation
decision = guard.evaluate({
    "action": "payment.create",
    "amount": 1200.0,
    "currency": "INR",
    "recipient": "merchant@store.com",
    "description": "Autonomous grocery restocking order",
    "prompt_context": "User prompt: Restock organic coffee beans.",
})

if decision.status == "ALLOW":
    print(f"Action permitted! Policy version: {decision.policy_version}")
    # Safe to initiate Razorpay order
elif decision.status == "HOLD":
    print(f"Action held for human review: {decision.pending_approval_id}")
    # Wait for operator in AgentGuard dashboard
else:
    print(f"Blocked by AgentGuard: {decision.reason}")
    for violation in decision.violations:
        print(f" - {violation}")

Context Manager Usage

with AgentGuard(base_url="http://localhost:3001", api_key="ag_live_...") as guard:
    decision = guard.evaluate(action="payment.create", amount=25.0)
    if decision.is_allowed:
        pass

Strict Fail-Closed Guarantee

# evaluate_safely catches all timeouts, DNS failures, and gateway crashes,
# strictly returning an authoritative BLOCK decision so your code never fails open.
decision = guard.evaluate_safely(action="payment.create", amount=50.0)
if not decision.is_allowed:
    print(f"Transaction denied: {decision.reason}")

License

MIT © AgentGuard Security Team