描述
Preview
Give your NPCs a real brain — offline.
LocalMind AI runs a genuine large language model inside your game, on the player's machine. No cloud service, no API keys, no internet requirement, and no cost per message. Ship it, and every copy of your game has an AI brain built in.
The model never loads into your game's process. It runs in a separate helper process over a private local pipe — so a model crash, an out-of-memory spike, or a bad GGUF file kills the helper, not your title. Your game gets an error event and carries on, and the plugin restarts the helper automatically.
Key Features
🧠 Fully offline inference — no internet, no API keys, no per-token cost, ever
🛡️ Crash-isolated — the model runs out-of-process with auto-restart and recovery events
💬 Real conversations — Sessions with history, streaming tokens, and separated chain-of-thought
👁️ Vision — send the viewport, a Texture2D, a render target, or an image file and the NPC responds to what's actually there
🎙️ Voice both ways — push-to-talk mic capture in, spatialised text-to-speech out
🔧 Tool calling — the model triggers real gameplay functions: hand over an item, open a door, start a quest
📐 Grammar-constrained output — guaranteed-parseable JSON decisions, invalid replies are structurally impossible
🧩 Semantic memory — NPCs recall facts by meaning, across sessions and save games
🎭 Persona Data Assets — designers write characters in the Content Browser, no Blueprint, no recompiles
📦 Built-in Model Manager — hardware detection, curated catalog, Hugging Face search, "will it fit?", and test chat in the editor
⚡ Frame-rate safe — GPU yield, token rate limiting, below-normal priority, second-GPU targeting
🧵 Blueprint & C++ — everything async, nothing blocks your game thread
Perfect for
Talking NPCs · dynamic quest and dialogue systems · AI companions that see the world · shopkeepers and quest-givers that actually act · behaviour-tree decision making · any game that wants AI without a cloud bill.
How it works
Isolate — the plugin launches your own executable as a hidden worker that loads the inference helper. No standalone .exe is shipped.
Load — one GGUF model sits in that process. Many Sessions share it; opening a session is cheap, so every NPC gets its own.
Generate — tokens stream back over the pipe to Blueprint events. Constrain them with a grammar, or let the model call your tools.
Ship it your way
Package with build — models staged into your build; the editor keeps packaging settings in sync for you.
Download in-game — small installer, model fetched on first run. Full progress, speed, pause, and true resume via HTTP Range — no bytes re-downloaded.
Technical
Unreal Engine 5.3-5.7 · Windows 64-bit · C++ code plugin · two modules (Runtime + Editor)
Blueprint and C++ projects both fully supported
CPU and Vulkan backends ship prebuilt — you never compile the AI engine
MSVC runtime included; players need no redistributable
Example content: demo NPC, example Blueprint, Persona asset, and a working model-downloader widget
66-page illustrated manual included — every screenshot is a Blueprint graph you can rebuild node for node
Before you buy
No model is included — download one in-editor from the catalog or Hugging Face. Budget 0.4–20 GB depending on size; a 3B Q4_K_M (~1.9 GB) runs on most gaming PCs.
Model licences are separate from the plugin licence. Some GGUF weights forbid commercial use — check before you ship.
Visual Studio 2022 is needed for the first build (Unreal compiles the plugin's own modules once).
Text-to-speech is the one online feature and is off by default. Everything else stays fully offline.
Documentation: LocalMind
Contact me
Discord