The Live2D Desktop Assistant That Parses Tools From Markdown

What it does NagaAgent is a cross-platform desktop AI assistant (Python backend, Electron frontend) wrapped in a “two-dimensional” anime aesthetic. It chats via LLM, displays a Live2D avatar, handles voice I/O, manages a Neo4j knowledge-graph memory, runs MCP tools and an embedded OpenClaw agent, and even recognizes game screens to offer strategy tips or automate play through MAA plugins. It also ships a music player, browser automation, and its own community forum with a points/affection system.

The interesting bit Rather than using native function-calling APIs, the engine makes the LLM write tool requests inside markdown ```tool``` blocks. A two-phase parser extracts these with json5 fault tolerance, then routes them in parallel to MCP servers, an OpenClaw agent, or Live2D animation triggers. This lets it work with any OpenAI-compatible provider that lacks tool-call support, and it will loop up to five rounds to complete a task.

Key highlights

• Markdown-based tool calls with json5 parsing, plus native Anthropic API format support via an api_format toggle
• “Memory Sea of Clouds”: 3D visualization of a Neo4j knowledge graph that feeds conversation context
• Game strategy mode with screen recognition and optional MAA automation integration
• Dual license: AGPL-3.0 or proprietary closed-source with written authorization
• Tight coupling to “NagaBusiness” cloud services (search, ASR, remote memory) when logged in, with local/Brave fallbacks when not

Caveats

• The repository description calls it a “simple agent framework,” but the README describes a monolithic, opinionated end-user application with gamification and a built-in marketplace
• README is Chinese-first; English and Japanese translations exist, yet the ecosystem (QQ bot linkage, Bilibili contact) is Sino-centric
• Requires Python 3.11 exactly (>=3.11, <3.12)

Verdict Worth a look if you want a batteries-included, anime-themed local AI companion with tool use and game automation. Not the right fit if you need a minimal, embeddable agent library for your own product.