RTGS2017/NagaAgent

A simple yet powerful agent framework for personal assistants, designed to enable intelligent interaction, multi-agent collaboration, and seamless tool integration.

What it solves

NagaAgent is a comprehensive AI assistant designed for the "anime-style" (ACG) community, combining a virtual Live2D avatar with advanced LLM capabilities. It solves the problem of creating a highly interactive, personalized desktop companion that can not only chat but also perform system tasks, manage long-term memories via knowledge graphs, and integrate with external tools and community forums.

How it works

The project uses a microservices architecture consisting of an API Server, Agent Server, MCP Server, and Voice Service, all coordinated by a central entry point.

  • Interaction: It utilizes a Vue 3/Electron frontend with a Live2D avatar that synchronizes mouth movements with TTS (Text-to-Speech) and follows the mouse cursor.
  • Intelligence: It supports various LLM providers (OpenAI, DeepSeek, Anthropic) and implements a streaming tool-calling loop that allows the AI to execute functions (like web searching or app launching) and feed the results back into the conversation.
  • Memory: It employs Graph-RAG (GRAG) to extract "quintuples" (subject, type, predicate, object, type) from conversations and store them in a Neo4j graph database for long-term retrieval.
  • Control: It integrates with OpenClaw for local computer control and the Model Context Protocol (MCP) for a pluggable architecture of specialized agents (e.g., weather, vision, IoT control).

Who it’s for

  • Users who want a visually engaging, anime-style AI companion on their desktop.
  • Power users looking for an agentic AI that can control their local system and browse the web.
  • Developers interested in integrating Live2D avatars with LLMs and Graph-RAG memory systems.

Highlights

  • Live2D Integration: High-fidelity virtual avatar with a 4-channel orthogonal animation system and SSAA anti-aliasing.
  • Graph-RAG Memory: Automatic extraction of facts into a 3D visualized "Mind Sea" using Neo4j.
  • Pluggable Tools: Support for MCP (Model Context Protocol) and community-installed skills.
  • Multimodal Interaction: Full-duplex real-time voice interaction (via Qwen Omni), ASR, and screen capture for vision-based tasks.
  • Context Compression: Automatic summarization of long conversations (over 100k tokens) to prevent context overflow.