Local AI Chat

New chat

Settings

Model

Custom model ID

Overrides the dropdown. Must be a public ONNX model compatible with Transformers.js.

Quantization (dtype)

Context window –

Memory estimate –

Hardware

Compatibility tests

Loads each model once, runs a small inference, and records what works on this GPU. Several minutes on first run.

Try every supported dtype (slower)

Text to speech

Kokoro runs locally in your browser. Replies are spoken only when you press Play.

Kokoro voice

Speed 1.00

Generation

Temperature 0.70

Lower = focused, higher = creative.

Top-p 0.95

Nucleus sampling cutoff.

Repetition penalty 1.10

Max new tokens 192

16 1024

Lower is faster and uses less memory.

Prompt context tokens 1,536

256 8192

Recent chat history sent to the model. Lower is faster.

Web search

When enabled, every prompt uses search context automatically.

SearxNG URL

Optional. Uses SearxNG first, then DuckDuckGo Instant Answer and Wikipedia when needed.

MCP bridge

Local companion bridge used by the static app to reach MCP servers.

MCP server URLs

One Streamable HTTP or SSE MCP URL per line. When any URL is saved, MCP tools are active for every prompt.

WebGPU setup

Chrome / Edge

Open chrome://flags.
Enable Unsafe WebGPU.
Relaunch the browser.

Linux / AMD

Open chrome://flags.
Enable Vulkan.
Open chrome://gpu and confirm WebGPU is hardware accelerated.
In this app, Settings → Hardware should show the real GPU, not google · swiftshader.

White video after Vulkan

In chrome://flags, set Choose ANGLE graphics backend to OpenGL.
If video is still white, set Zero-copy rasterizer to Disabled.
If needed, disable only Hardware-accelerated video decode.

Keep Chrome Settings → System → Use graphics acceleration when available turned on.

Chrome may block direct chrome:// links from web pages. If a link does not open, copy it into the address bar.

Downloaded models

Scanning browser cache…

Loading model

This may take a few minutes on first load.

Initializing…

Welcome to Local AI

Load a model to start chatting. Everything runs on your device.

Context

Runs locally via WebGPU. Chats stored in your browser.