Claude Agent SDK Integration

Strathon enforces policies on Claude Agent SDK tool calls before they execute, with the full action set including interactive approval. Enforcement uses the SDK's first-class PreToolUse/PostToolUse hooks on ClaudeSDKClient; a query() wrapper adds observability for code not using the client.

Enforcement scope: full on ClaudeSDKClient, where the async hooks enforce all seven actions, including interactive require_approval. The module-level query() function does not support hooks, so that path is observability-only.

Installation

bash

pip install "strathon[claude-agent]"

Setup

python

from strathon import Client, instrument

client = Client(
    api_key="stra_...",
    endpoint="http://localhost:4318",
)
instrument(client, frameworks=["claude_agent"])

What Gets Captured

Agent queries: input, model, parameters
Tool use: tool name, arguments, results
Agent reasoning: thinking content (if extended thinking enabled)
Token usage: input tokens, output tokens
Latency: per-query timing

Example Policy

Block tool calls that access sensitive directories:

cel

attrs["gen_ai.tool.name"] == "file_read"
  && attrs["strathon.tool.args"].contains("/etc/")

Require approval for agent actions that modify infrastructure:

cel

attrs["gen_ai.tool.name"] in ["deploy", "rollback", "scale"]

Notes

Wraps query() for tracing and policy enforcement.
Captures extended thinking content as span attributes.
Requires claude-agent-sdk>=0.1.0 (installed by the claude-agent extra).