Getting started

NobodyWho lets you run large language models inside your Godot game, entirely offline. Chat with NPCs, transcribe player speech, generate voice lines, and embed text for search — all on the player's own machine, with no servers, no API keys, and no telemetry.

How do I get started?

NobodyWho is a GDExtension. Install it by copying the addons/nobodywho folder from a release into your project:

your-project/
└── addons/
    └── nobodywho/
        ├── nobodywho.gdextension
        ├── libnobodywho-godot-x86_64-unknown-linux-gnu-release.so
        └── ...

The release zip bundles the libraries for all supported platforms (Windows, Linux, macOS, and Android), so the same folder works everywhere. Restart the editor after copying it in, and the NobodyWho classes become available to GDScript.

info

NobodyWho requires Godot 4.6 or newer.

With the extension installed, your first conversation is three lines:

func _ready():
    var chat = await NobodyWhoChat.create("hf://NobodyWho/Qwen_Qwen3-0.6B-GGUF", {})
    var stream = chat.ask("Hello! Who are you?")
    print(await stream.completed())

The model downloads to a local cache on first use — after that it works completely offline. See Downloading models for the supported path formats, and Chat for the full chat walkthrough.

warning

Every async method resolves to the result, or null on failure (with details in the editor's Errors tab). Always await the returned value immediately — storing it and awaiting it later is unsupported and may hang.

Logging

The library logs through Godot's console (godot_print/godot_warn/godot_error). By default it logs at INFO level; make it chattier while debugging with:

NobodyWho.set_log_level("DEBUG")

Valid levels are "TRACE", "DEBUG", "INFO", "WARN", and "ERROR".

What's available?

• Chat — talk to local LLMs, with full control over the context.
• Tool calling — let the model call your GDScript functions.
• Multimodal models — image and audio input.
• Sampling — tune token selection and constrain output to a format.
• Speech to text — Whisper transcription from files or microphone.
• Text to speech — generate voice lines as WAV bytes.
• Voice activity detection — know when the player starts and stops speaking.
• Embeddings & RAG — semantic search over your game's text.

Minimum recommended specs

• CPU: 4 cores
• RAM: 8 GB
• GPU: optional (Vulkan) — speeds up all models significantly

Smaller models (around 0.5-1B parameters, quantized) run comfortably on these specs, including on phones. Every feature works without a GPU, just slower.

Feedback & Contributions

Bugs and feature requests go to our issues. Want to chat or need help? Join us on Discord. Contributions are welcome — see CONTRIBUTING for how to build from source.