sasy

Python client for SASY. It records the inputs behind an agent's actions and asks the policy engine for a decision before each protected tool runs.

The engine runs separately; follow local engine setup before running an agent.

Install

Install in a Python 3.11 or newer environment, then start the local engine with Docker running:

pip install sasy
sasy engine start

For repository examples, use the checkout setup.

Protect an agent

Local setup saves a selected connection profile in ~/.sasy. The Python SDK uses it from any project unless you configure another endpoint. Enable instrumentation before building any agents:

import sasy

sasy.configure()
sasy.instrument()

Run each task inside with sasy.session(policy=...). The session records its messages and binds the policy used to check tool calls. Cooperating agents share the same session.

Supported frameworks: LangChain, Google ADK, and Langroid.

For a complete runnable program, follow framework examples or try the LangChain examples.

Instrumentation enables supported installed frameworks. An unusable framework is skipped with a warning; pass its flag as True, such as sasy.instrument(langchain=True), to require it. HTTP checks are off by default. See instrumentation options for selection and HTTP behavior.

Use the client directly

import json
from pathlib import Path

import sasy

sasy.configure()
with sasy.session(policy=Path("read-only.dl")):
    decision = sasy.check_tool_call("Read", json.dumps({"file_path": "notes.txt"}))
    if decision.authorized:
        # Perform the protected action here.
        print(Path("notes.txt").read_text())

Supply your own policy file. For checks that depend on earlier messages, your integration must also record those dependencies. See the adapter guide.

See configuration, authentication, and the security model for setup and supported boundaries.