fangcun-hook-sdk 0.1.8

Inline-hook client SDK for FangcunGuard. Lets agents call the FangcunGuard Hook server for input/output safety scanning without changing their LLM URL or API key.

fangcun-hook-sdk

Inline-hook client SDK for FangcunGuard. Lets your agent call a FangcunGuard Hook server for input/output safety scanning without changing the agent's LLM URL or API key.

agent (your code, your LLM key, your LLM URL)
   │
   ├─ pre-LLM   ──► hook.scan_input(messages)   ──► allow / block / replace / anonymize
   ├─ call your LLM as usual
   ├─ post-LLM  ──► hook.scan_output(content)   ──► allow / block / replace / restore
   ▼
agent receives the final message

Install

pip install fangcun-hook-sdk

# optional: bring the framework you actually use
pip install "fangcun-hook-sdk[openai]"   # raw OpenAI client wrapper
pip install "fangcun-hook-sdk[agents]"   # OpenAI Agents SDK guardrails

Minimal example (raw OpenAI client)

from openai import OpenAI
from fangcun_hook_sdk import HookClient
from fangcun_hook_sdk.adapters.openai_raw import wrap_openai

# Your existing LLM client — unchanged.
llm = OpenAI(api_key="sk-...", base_url="https://api.openai.com/v1")

# Wrap it once. The agent code below sees no difference.
hook = HookClient(base_url="http://localhost:5002", api_key="sk-xxai-...")
llm = wrap_openai(llm, hook)

resp = llm.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "My email is [e-mail]"}],
)
print(resp.choices[0].message.content)
# PII is anonymized before reaching the LLM and restored on the way back.

Minimal example (OpenAI Agents SDK)

from agents import Agent, Runner
from fangcun_hook_sdk import AsyncHookClient
from fangcun_hook_sdk.adapters.openai_agents import (
    make_input_guardrail, make_output_guardrail,
)

hook = AsyncHookClient(base_url="http://localhost:5002", api_key="sk-xxai-...")

agent = Agent(
    name="support-bot",
    instructions="You are a helpful assistant.",
    input_guardrails=[make_input_guardrail(hook)],
    output_guardrails=[make_output_guardrail(hook)],
)

result = await Runner.run(agent, "What is the weather in Tokyo?")

Low-level API

If your framework isn't covered, call the client directly:

from fangcun_hook_sdk import HookClient, Action

hook = HookClient(base_url="http://localhost:5002", api_key="sk-xxai-...")

scan_in = hook.scan_input(messages=[{"role": "user", "content": "..."}])
if scan_in.action == Action.BLOCK:
    raise RuntimeError(scan_in.message)
elif scan_in.action == Action.ANONYMIZE:
    messages_to_send = scan_in.anonymized_messages
    restore_mapping  = scan_in.restore_mapping
else:
    messages_to_send = original_messages
    restore_mapping  = None

# ... call your LLM with messages_to_send ...

scan_out = hook.scan_output(
    content=llm_reply,
    restore_mapping=restore_mapping,
)
if scan_out.action == Action.BLOCK:
    raise RuntimeError(scan_out.message)
final_text = scan_out.content if scan_out.action == Action.RESTORE else llm_reply

Failure mode

By default the SDK is fail-open: if the hook server is unreachable or returns an error, calls return Action.PASS and the agent keeps working. Pass fail_open=False to surface those errors as exceptions instead.

hook = HookClient(..., fail_open=False)

Talking to a legacy server

If you're pointing at an older FangcunGuard server that only exposes /v1/gateway/process-input and /v1/gateway/process-output, set:

hook = HookClient(..., primary_path="/v1/gateway")

The new server registers both prefixes, so most users don't need this.