AgentShield Python SDK

Lightweight security guardrails for AI agents. All security logic runs server-side.

Quick Start

from agentshield import Shield

shield = Shield()  # reads AGENTSHIELD_API_KEY from env

@shield.guard
async def send_email(to: str, body: str) -> str:
    ...  # your tool implementation

# The server decides: ALLOW, BLOCK, or REQUIRE_CONFIRMATION
await send_email(to="user@company.com", body="Hello")

Session Mode

async with shield.session("Summarize my emails") as s:
    result = await s.guarded_executor.execute(
        "read_inbox", {"limit": 10}, read_inbox_fn
    )

Error Handling

from agentshield import Shield, ToolCallBlocked, ConfirmationRejected, ServerError

shield = Shield()

@shield.guard
async def send_email(to: str, body: str) -> str:
    ...

try:
    await send_email(to="user@test.com", body="hi")
except ToolCallBlocked as e:
    print(f"Blocked: {e.reason} (trace: {e.trace_id})")
except ConfirmationRejected:
    print("User declined confirmation")
except ServerError as e:
    print(f"Server error: {e}")

Configuration

# Explicit configuration
shield = Shield(
    api_key="your-key",
    base_url="https://shield.yourcompany.com",
    timeout=10.0,
    max_retries=3,
    agent_id="my-agent",
)

Or via environment variables:

• AGENTSHIELD_API_KEY (required)
• AGENTSHIELD_BASE_URL (default: http://localhost:8000)
• AGENTSHIELD_TIMEOUT (default: 10.0)
• AGENTSHIELD_AGENT_ID

Or via agentshield.yaml in the working directory.

Data Sanitization

# Sanitize external data before processing
result = await shield.sanitize(
    data=email_body,
    source="email/external",
)
# result.content has hidden injections removed
# result.trust_level shows the computed trust level

Framework Integrations

from agentshield.integrations import LangChainShield, CrewAIShield

# LangChain
guarded = LangChainShield(shield).wrap(agent_executor)

# CrewAI
guarded = CrewAIShield(shield).wrap(crew)

# AutoGen
from agentshield.integrations import AutoGenShield
AutoGenShield(shield).wrap(assistant)

# Claude Agent SDK
from agentshield.integrations import ClaudeAgentShield
guarded_handler = ClaudeAgentShield(shield).wrap(my_tool_handler)