Claude Agents Integration

Overview

ClaudeAgentGuard wraps an Anthropic client so every messages.create() response is validated before it reaches your code or triggers a tool call.

Tier: Team+

Why It Matters

Claude Agents run autonomous tool-use loops. Governance must intercept loop outputs without breaking the agent's tool execution flow.

Prerequisites

pip install agentrust-py
# Team+ API key
# Claude Agents SDK installed in your environment

Step-by-Step Guide

wrap_client() returns a proxy — it is not a context manager. Use the returned object in place of the original client.

from anthropic import Anthropic
from agentrust_sdk import ClaudeAgentGuard

guard = ClaudeAgentGuard(
    agent_id="claude-agent",
    base_url="http://127.0.0.1:8765",   # adapter default is :8000
)

client = Anthropic()
governed = guard.wrap_client(client)     # use `governed` from here on

response = governed.messages.create(
    model="claude-opus-4-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Analyze this document"}],
)

To govern a whole agent loop instead of individual calls, wrap the loop function:

@guard.wrap_loop
def run_agent(query: str) -> str:
    ...

Each governed call sends the response text and any tool calls to the gateway; with the default block_on_block=True a block decision raises BlockedError.

Examples

See agentrust_sdk/agentrust_sdk/adapters/claude_agents.py in the repository for implementation details.

Best Practices

• Wrap at the client level for consistent coverage
• Use block_on_block=True for sensitive document processing
• Combine with PII policy pack for customer data

Common Mistakes

• Catching TierGateError — the guard raises plain RuntimeError at construction
• Calling guard.wrap(...) — the methods are wrap_client() and wrap_loop()
• Continuing to use the original client instead of the returned proxy

Troubleshooting

Related Resources

• OpenAI Agents
• Tiers and Capabilities