strathon 1.3.0

Open-source AI agent firewall. Blocks dangerous agent tool calls before they execute, with in-process CEL enforcement at the tool-call boundary.

Strathon SDK

Python SDK for Strathon, the open-source AI agent firewall.

Install

pip install strathon

Optional framework integrations:

pip install "strathon[openai]"          # OpenAI
pip install "strathon[openai-agents]"   # OpenAI Agents SDK
pip install "strathon[anthropic]"       # Anthropic
pip install "strathon[claude-agent]"    # Claude Agent SDK
pip install "strathon[langgraph]"       # LangGraph
pip install "strathon[langchain]"       # LangChain
pip install "strathon[crewai]"          # CrewAI
pip install "strathon[autogen]"         # AutoGen
pip install "strathon[pydantic-ai]"     # Pydantic AI
pip install "strathon[google-adk]"      # Google ADK
pip install "strathon[all]"             # All 10 frameworks

Quick start

from strathon import Client, instrument

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

# Your agent code runs as normal. Strathon traces every call
# and enforces CEL policies before tool execution.

Enforce

A policy with action block stops the call before the tool body runs:

from strathon import StrathonPolicyBlocked

try:
    agent.run("Email the Q3 numbers to [e-mail]")
except StrathonPolicyBlocked as e:
    print(f"Blocked by policy: {e.policy_name}")
    # The tool call never executed. Recorded in the audit trail.

Seven actions: block, steer, throttle, log, alert, require_approval, allow.

What gets captured

• Every LLM call: model, tokens in/out, cost, latency
• Tool calls: name, arguments, result, duration
• Agent topology: parent-child relationships, handoffs
• Errors: failed tool calls, model errors, timeouts

Documentation

• Quickstart
• Framework guides
• CEL policy reference
• GitHub

License

Apache License 2.0. See LICENSE.