agr-sdk

Python SDK for AGR (Agentic Governance Runtime) — a drop-in governance layer for AI agent tool calls. Every action your agent takes is evaluated against your org's policies before it runs.

Install

pip install agr-sdk

Quickstart — zero-touch guard

import agr

agr.configure(
    base_url="https://api.agr.dev",   # or your self-hosted AGR instance
    api_key="agr_sk_xxx",
    agent_id="my-agent",
)

@agr.guard(action="execute_sql", resource="prod_db")
def run_query(sql: str) -> str:
    ...

Calling run_query(...) now evaluates the action against your AGR policies first. If policy denies it, AGRDeniedError is raised instead of the query running. If policy requires human approval, the call blocks (or raises AGRApprovalTimeoutError) until a human decides.

Quickstart — wrap existing tools

import agr

agr.configure(base_url="https://api.agr.dev", api_key="agr_sk_xxx", agent_id="my-agent")
tools = agr.wrap_tools(my_langchain_tools)  # LangChain BaseTool instances, plain callables, or async callables

Quickstart — raw client

from agr import AGRClient

client = AGRClient(api_key="agr_sk_xxx", base_url="https://api.agr.dev")
result = client.evaluate(agent="my-agent", action="deploy", resource="production")

if result.allowed:
    ...
elif result.denied:
    ...
elif result.requires_approval:
    ...

An AsyncAGRClient with the same shape is available for async code.

Framework integrations

• LangChain — agr.wrap_tools() auto-detects BaseTool instances.
• CrewAI / LangGraph — see agr.plugins.crewai / agr.plugins.langgraph.

License

MIT