agentguard-tech 0.1.0

Runtime security for AI agents — policy engine, audit trail, and kill switch

agentguard

Runtime security for AI agents — policy engine, audit trail, and kill switch.

Overview

AgentGuard gives AI agents production-grade guardrails:

• 🛡️ Policy evaluation — check every tool call before execution
• 📋 Audit trail — tamper-evident hash chain of every action
• 🔴 Kill switch — instantly halt all agents
• 🔍 Audit verification — cryptographically verify the audit chain
• ⚡ Zero dependencies — pure Python stdlib, works anywhere

---

Installation

pip install agentguard

Requires Python 3.8+. No external dependencies.

---

Quick Start

from agentguard import AgentGuard

guard = AgentGuard(api_key="ag_your_api_key")

# Evaluate an agent action before executing it
decision = guard.evaluate(
    tool="send_email",
    params={"to": "[e-mail]", "subject": "Hello"}
)

if decision["result"] == "allow":
    print("Action allowed, risk score:", decision["riskScore"])
    # proceed with tool execution
elif decision["result"] == "block":
    print("Action blocked:", decision["reason"])
elif decision["result"] == "require_approval":
    print("Waiting for human approval...")
elif decision["result"] == "monitor":
    print("Action monitored (allowed but logged):", decision["reason"])

---

API Reference

AgentGuard(api_key, base_url=...)

Create a client instance.

guard = AgentGuard(
    api_key="ag_your_api_key",
    base_url="https://api.agentguard.tech"  # optional, default shown
)

---

evaluate(tool, params=None) → dict

Evaluate a tool call against your policy. Call this before every tool execution.

decision = guard.evaluate("read_file", {"path": "/data/report.csv"})
# Returns:
# {
#   "result": "allow",          # allow | block | monitor | require_approval
#   "riskScore": 5,             # 0-1000
#   "reason": "Matched allow-read rule",
#   "durationMs": 1.2,
#   "matchedRuleId": "allow-read"  # optional
# }

Integration pattern:

def safe_tool_call(tool_name, tool_func, **params):
    decision = guard.evaluate(tool_name, params)
    if decision["result"] in ("allow", "monitor"):
        return tool_func(**params)
    elif decision["result"] == "block":
        raise PermissionError(f"Blocked by policy: {decision['reason']}")
    elif decision["result"] == "require_approval":
        raise PermissionError("Awaiting human approval")

---

get_usage() → dict

Get usage statistics for your tenant.

usage = guard.get_usage()
print(usage)
# {
#   "requestsToday": 142,
#   "requestsThisMonth": 3891,
#   "plan": "pro",
#   "limits": { "requestsPerDay": 10000 }
# }

---

get_audit(limit=50, offset=0) → dict

Get audit trail events with pagination.

audit = guard.get_audit(limit=100, offset=0)
for event in audit["events"]:
    print(f"{event['timestamp']} | {event['tool']} | {event['decision']}")

---

kill_switch(active) → dict

Activate or deactivate the global kill switch.

# Emergency halt — stop all agents immediately
guard.kill_switch(True)

# Resume operations
guard.kill_switch(False)

---

verify_audit() → dict

Verify the cryptographic integrity of the audit hash chain.

result = guard.verify_audit()
if result["valid"]:
    print("Audit chain is intact")
else:
    print(f"Chain broken at event index: {result['invalidAt']}")

---

Complete Example — LangChain-style Agent

from agentguard import AgentGuard

guard = AgentGuard(api_key="ag_your_api_key")

def run_tool(name: str, func, **params):
    """Execute a tool with AgentGuard policy enforcement."""
    decision = guard.evaluate(name, params)
    
    result = decision["result"]
    if result == "block":
        raise PermissionError(f"Policy blocked {name}: {decision['reason']}")
    if result == "require_approval":
        raise PermissionError(f"Human approval required for {name}")
    
    # "allow" or "monitor" — proceed
    return func(**params)


# Your tools
def send_email(to: str, subject: str, body: str) -> str:
    # ... send the email
    return f"Email sent to {to}"

def read_file(path: str) -> str:
    with open(path) as f:
        return f.read()


# Use with policy enforcement
content = run_tool("read_file", read_file, path="/data/report.csv")
run_tool("send_email", send_email, to="[e-mail]", subject="Report", body=content)

---

Error Handling

from agentguard import AgentGuard

guard = AgentGuard(api_key="ag_your_key")

try:
    decision = guard.evaluate("dangerous_tool", {"target": "production_db"})
except RuntimeError as e:
    print(f"API error: {e}")
    # RuntimeError: AgentGuard API error: 401 Unauthorized

---

Links

• 🌐 agentguard.tech
• 🎮 Live Demo
• 📦 GitHub
• 📘 npm SDK

License

MIT