CoreAI — open-source agent framework for Unity built on top of LLMUnity (function calling, tools, memory, chat UI)

[NeoXider]

Hi! I've been building games with LLMUnity as the local backend and ended up extracting my agent layer into an open-source framework: CoreAI — sharing it here in case it's useful to others running local models.

What it adds on top of LLMUnity:

• Function calling into your game code — the model calls C# methods you register as tools (spawn objects, move NPCs, query scene state), with automatic retry/self-repair when the model gets a tool call wrong.
• Agents in ~3 lines — AgentBuilder with roles, per-role model routing, and a one-liner CoreAi.AskAsync("...") for quick starts.
• Persistent agent memory + conversation summarization with a token budget for long chats.
• Drop-in chat UI — menu item creates a working streaming chat scene (token-by-token).
• Optional Lua sandbox — the model can write gameplay scripts executed safely as coroutines.
• Works with LLMUnity (local GGUF) or any OpenAI-compatible API — same code either way. Everything is tested on a 4 GB local model; the test matrix by model size is in the README.

LLMUnity is auto-detected as an optional module, so CoreAI compiles with or without it — but local-first via LLMUnity is the primary path and what I use myself.

Repo (MIT-adjacent PolyForm license, Unity 6000+): https://github.com/NeoXider/CoreAI

Huge thanks for LLMUnity — it's what made shipping local LLM gameplay realistic. Feedback and issues welcome!