litellm-adk 0.2.3

Production-grade multiservice Agent Development Kit

🤖 LiteLLM ADK (Agent Development Kit)

A production-grade, highly flexible multiservice framework for building AI agents. Swap models, API keys, and configurations dynamically while maintaining a robust structure for tool usage, complex memory persistence, and human oversight.

---

🌟 Key Features

• 🔌 Model Agnostic: Native support for 100+ LLMs (OpenAI, Anthropic, OCI Grok-3, Llama) via LiteLLM.
• 📊 Database Intelligence: Built-in DatabaseAgent (NL2SQL) with adapters for MySQL, PostgreSQL, and MongoDB.
• 🧠 Advanced Memory: Pluggable persistence via local JSON, MongoDB, or Vector stores (Chroma, PGVector).
• 🛡️ Human-in-the-Loop: Integrated PolicyEngine and ApprovalManager for safe tool execution with audit trails.
• ⚡ Modern Async/Streaming: First-class support for ainvoke and event-based streaming (astream).
• 🔍 Observability: Structured logging and token tracking out of the box.

---

🚀 Quick Start

1. Installation

pip install litellm-adk

2. Basic Agent with Tools

import asyncio
from litellm_adk import LiteLLMAgent, tool, FileMemory

# Define a tool with zero boilerplate
@tool
def get_stock_price(symbol: str):
    """Fetches the current stock price."""
    return {"symbol": symbol, "price": 150.0}

async def main():
    # Initialize with persistent file memory
    agent = LiteLLMAgent(
        model="oci/xai.grok-3",
        tools=[get_stock_price],
        memory=FileMemory("history.json")
    )
    
    response = await agent.ainvoke("What is the price of AAPL?")
    print(f"Agent: {response}")
    await agent.aclose()

if __name__ == "__main__":
    asyncio.run(main())

---

📊 High-Performance Database Agents

The ADK specializes in NL2SQL via the DatabaseAgent. It supports "Blind Summarization" for security—allowing the agent to talk about data without leaking PII into model logs.

from litellm_adk import DatabaseAgent

agent = DatabaseAgent(
    db_url="mysql+pymysql://user:pass@localhost/analytics",
    model="gpt-4o",
    return_direct=True # Privacy: Summary to LLM, full data to your UI
)

---

📖 Documentation

Dive deeper into the components:

• 🤖 Agent Core: Logic, streaming, and failover.
• 📊 Database Integration: SQL, MySQL, and MongoDB.
• 🧠 Memory & Persistence: Storing state and history.
• 🛠️ Tools & HITL: Custom tools and Human-in-the-Loop policies.

---

🏗️ Project Structure

• core/: Abstract base classes (Agent, Memory, Adapter).
• integrations/: Specific implementations (MySQL, MongoDB, ChromaDB).
• session/: State management for multi-user chat.
• observability/: Logging and telemetry.

---

📜 License

MIT © 2026 Aarumugapandi