sigui-sdk 0.2.0

Python SDK for Sigui Protocol — Autonomous Security for the Agentic Economy

Sigui SDK — Autonomous Security for AI Agents

Sigui is an open-source security oracle that protects AI agents from sending erroneous or malicious crypto payments.

Add a security evaluation layer to any AI agent in 3 lines of code. Every transaction is analyzed by a 5-layer AI security pipeline (including a fine-tuned Vision Language Model) in < 50 ms before execution.

---

⚡ The Problem & The Solution

The Problem: Autonomous agents (LangChain, CrewAI, AutoGen) can now execute USDC transfers and interact with DeFi protocols. But a single hallucination, prompt injection, or misconfigured tool means the agent can send $5,000 to the wrong address. There is no middleware to stop it.

The Solution: sigui-sdk intercepts the payment intent, evaluates it against a 5-layer security pipeline, and returns a strict verdict (ALLOW, BLOCK, ESCALATE).

---

📦 Installation

pip install sigui-sdk

Install the specific extras for the framework you are using:

pip install "sigui-sdk[langchain]"       # LangChain & LangGraph
pip install "sigui-sdk[crewai]"          # CrewAI
pip install "sigui-sdk[autogen]"         # Microsoft AutoGen (AG2)
pip install "sigui-sdk[openai-agents]"   # OpenAI Agents SDK
pip install "sigui-sdk[smolagents]"      # HuggingFace smolagents

pip install "sigui-sdk[all]"             # Install all integrations

---

🚀 Quickstart (2 lines)

from sigui import SiguiClient

async with SiguiClient(api_url="http://localhost:8000") as client:
    result = await client.evaluate(amount=5.0, destination="0xRecipient...")

    if result.is_safe:
        print(f"✅ Authorized   risk={result.risk_score:.3f}")
    elif result.is_blocked:
        print(f"🚫 Blocked      {result.reason}")
    else:
        print(f"⚠️  Escalation required")

(Note: Sigui requires the backend security engine to be running locally. See the Backend Setup section below).

---

🧩 Framework Integrations

Sigui provides native Tools for all major agent frameworks. Drop them into your agent's tool list with zero refactoring.

🦜 LangChain / LangGraph

from sigui import SiguiClient
from sigui.integrations.langchain import create_langchain_tool

client = SiguiClient(api_url="http://localhost:8000")
sigui_tool = create_langchain_tool(client, auto_escalate=True)

# Drop into any LangChain agent
agent = initialize_agent(tools=[sigui_tool, ...], llm=llm)

🤖 CrewAI

from sigui import SiguiClientSync
from sigui.integrations.crewai import SiguiEvaluationTool

client = SiguiClientSync(api_url="http://localhost:8000", agent_id="my_crew")
tool = SiguiEvaluationTool(sigui_client=client, auto_escalate=True)

payment_agent = Agent(role="DeFi Agent", tools=[tool], ...)

🧩 Microsoft AutoGen (AG2)

from autogen_agentchat.agents import AssistantAgent
from sigui import SiguiClient
from sigui.integrations.autogen import create_autogen_tool

client = SiguiClient(api_url="http://localhost:8000")
sigui_tool = create_autogen_tool(client, auto_escalate=True)

agent = AssistantAgent(name="payment_agent", tools=[sigui_tool], ...)

🤗 HuggingFace smolagents

from smolagents import CodeAgent
from sigui import SiguiClient
from sigui.integrations.smolagents import SiguiTool

client = SiguiClient(api_url="http://localhost:8000")
tool = SiguiTool(client, auto_escalate=True)

agent = CodeAgent(tools=[tool], model=...)

---

🛡️ Framework-Agnostic Decorator

If you don't want to use agent Tools, you can gate any async Python function directly:

from sigui.decorators import sigui_protect

client = SiguiClient(api_url="http://localhost:8000")

@sigui_protect(client, amount_arg="usdc", destination_arg="to")
async def transfer(to: str, usdc: float, memo: str = ""):
    # This code ONLY executes if Sigui returns ALLOW
    await wallet.send(to, usdc)

# Usage (raises SiguiBlockedError if flagged as malicious)
await transfer(to="0xAttacker...", usdc=500.0)

---

🧠 How it Works: The 5-Layer Pipeline

When your agent calls client.evaluate(), the backend runs a comprehensive security check in under 50ms:

1. MemoClaw: Episodic behavioral memory. (Is this agent acting weird compared to its history?)
2. Sirige: Rule-based anomaly detection (spikes in amounts, blacklisted addresses).
3. Anti-splitting: Cross-chain flow analysis to detect smurfing.
4. Imina Na (Vision): A fine-tuned Vision model that renders the transaction graph topology as an image and classifies the attack pattern.
5. Kanaga Risk Aggregator: Final risk scoring.

🤗 Open-Source Vision Model & Dataset

The Imina Na vision layer is powered by Qwen2-VL-7B-Instruct, fine-tuned using Unsloth on AMD MI300X GPUs. Both the model and the 1-million sample training dataset are open-source and available on HuggingFace:

• Model: Ibonon/imina_na_v2_lora
• Dataset: Ibonon/sigui-depin-1mn

---

⚙️ Backend Setup (Required)

sigui-sdk requires the Sigui Security Engine to be running. We do not currently provide a public hosted API.

You can run the backend locally in 2 minutes using Docker:

# 1. Clone the main repository
git clone https://github.com/ibonon/Sigui.git
cd Sigui

# 2. Configure for local dev (no crypto keys required)
cp .env.example .env
# Ensure DEMO_MODE=true is set in .env

# 3. Launch the security oracle
docker compose up

Once running, your SDK can connect to http://localhost:8000.

---

📄 License

MIT © Sigui Protocol.

Built for the Agentic Economy.