Local AI · experimental · v0.1.2

Your prompt.
Your GPU.

Run an assistant, plan an app or draft a small page in your browser. No API key or credits needed.

This local mode is separate from server-backed LibreChat agents. It has editable instructions, one task per run and no tools or wallet access.

1. Load a model

API cost: $0

Model

MB of model files, plus runtime downloads. Estimated GPU memory: about GB; actual needs vary. The first load can take several minutes.

This model needs shader-f16, which this adapter does not support. Choose Qwen 0.5B instead.

I agree to download this model and use this device's GPU, memory and power. Model files come from Hugging Face and GitHub and may be cached here. My prompt is processed locally.

Browser cannot access the GPU — what to check

In Chrome or Edge, enable “Use graphics acceleration when available” in Settings → System, restart the browser, then press Check GPU again. Inspect chrome://gpu to see the graphics driver and WebGPU status.

Chrome / Chromium on Linux: NVIDIA support is documented for Wayland. On our Linux Mint / X11 / NVIDIA test machine, a separate Chromium profile with Vulkan enabled worked:

chromium --user-data-dir=/tmp/lain-webgpu-test --enable-features=Vulkan --use-angle=vulkan https://lain-os.com/local

This is an experimental launch for a separate test profile; it does not change your usual browser settings. It is a tested workaround for that setup, not a guarantee for every Linux GPU or driver. A software adapter such as SwiftShader cannot run this GPU workspace.

Firefox on Linux: WebGPU is still experimental. Mozilla documents Linux support in Nightly; enabling a preference in a release build does not guarantee that a hardware adapter will be available. See Mozilla's WebGPU support notes.

Official WebGPU troubleshooting guide. A page cannot enable GPU access that the browser has disabled.

Keep this tab visible while loading or generating. Hiding or leaving it stops the work. Cached weights can be reused; closing the tab releases the model.

2. Choose an agent prompt

Starting point Agent name Instructions Your task

Output limit Result

Instructions and task together: up to 2800 UTF-8 bytes. Small models can make mistakes; keep tasks narrow.

Import JSON

3. Your result

The answer will stream here after you load a model and run a task. Nothing is sent to a hosted model.

Generating on this device…

Prompts and results are not saved to your Lain account or chat history. Export anything you want to keep before leaving.

A first step toward App Factory: portable agent prompts and local drafts. Deployment, tokens, DeFi modules and autonomous tools are not part of this experiment.

For developers: public v1 agent and model catalog. Execution runs in the user's browser, not through a server-side completion endpoint. Cyberia integration will use this boundary; this page does not accept third-party jobs yet.