MagiForge builds practical, on-device AI tools for Unreal Engine. Most "AI in games" plugins are thin wrappers around OpenAI or Anthropic endpoints — they need an API key, they bill per token, they leak network traffic from your shipped game, and they stop working the moment your players go offline. We took the other path. Everything we ship runs locally on the player's machine. Our flagship plugin MagiLLM brings real language model inference into the Unreal engine subsystem layer, powered by llama.cpp. You drop in any GGUF model — Gemma, Phi, Mistral, Llama 3.x — and stream tokens straight into your game thread through Blueprint-callable events (Say, OnToken, OnDone). Persona files are tag-based so you can ship multiple languages without recompiling. The plugin includes Gemma instruction-tuned weights out of the box, so chat works the moment the plugin is enabled — no setup, no API key, no cloud call. What we promise: • Source code included. Nothing closed-source that depends on UE headers. • Support for the three most recent major engine versions, kept current. • Documentation treated as a deliverable, not an afterthought — setup guide, API reference, persona authoring notes. • Bug reports get a real reply within a couple of business days. • Feature requests considered on the public roadmap. What we don't do: • Pretend the model output is reliable when it isn't. Small on-device models are good for casual NPC chatter and prototypes, not for legal advice or precision technical recall. • Charge per use. One purchase, no usage caps. If you're prototyping an AI feature for a game, an editor tool, or an offline assistant inside Unreal, and you'd rather not wire up a backend, we built this for you.