In-Browser AI Sandbox

Codetato

Run open-source AI models locally and 100% privately in your browser. Accelerated by WebGPU — no servers, no API keys, and no subscriptions.

Bandwidth Warning: Large File Downloads

Running models inside the browser requires downloading the model weights. The first time you load a model, your browser will download files ranging from 150MB to over 2.2GB.

  • Downloads will consume significant internet bandwidth. We recommend a flat-rate broadband or Wi-Fi connection.
  • Models are pulled directly from Hugging Face and automatically saved to your browser's local Cache Storage so they load instantly on subsequent visits (and work completely offline).

Current System Status

WebGPU Acceleration

Your browser or device does not support WebGPU (or it is disabled). Models will fallback to CPU (WASM), which is significantly slower.

Hardware Specifications

GPU DeviceDetecting...
Estimated VRAMDetecting...
CPU Logical CoresDetecting...

Model Registry & Expected Performance

The registry contains a list of models tested for compatibility. Note that these are provided as recommendations for in-browser use. Based on your detected hardware specifications, here are the expectations for initial model download sizes, compatibility status, and real-world execution speeds.

Model NameParams / QuantDownload SizeCompatibilityEst. Speed
Qwen 2.5 (0.5B)
General Chat
0.5Bq4~350 MBCPU only
WebGPU is unavailable. This model will run on your CPU via WASM, which is slower but usable.
Llama 3.2 (1B)
General Chat
1.2Bq4f16~750 MBVery slow
WebGPU is unavailable. This model size will be noticeably slow on CPU-only inference.
TinyLlama (1.1B)
General Chat
1.1Bq4~650 MBVery slow
WebGPU is unavailable. This model size will be noticeably slow on CPU-only inference.
Qwen 2.5 Coder (0.5B)
Code Generation
0.5Bq4~350 MBCPU only
WebGPU is unavailable. This model will run on your CPU via WASM, which is slower but usable.
Phi-3.5 Mini
Advanced Reasoning
3.8Bq4~2200 MBToo large for CPU
WebGPU is unavailable and this model is too large to run practically on CPU.

Development Roadmap

Follow our journey as we build the ultimate client-side, 100% private, browser-based AI workspace.

Completed

Local Models

Run open-source LLMs directly in your browser using WebGPU/WASM for 100% private inference.

Completed

System Theme

Automatic synchronization with light/dark OS modes and access to a robust curated theme selector.

In Progress

Agent Sandbox

A secure client-side sandbox environment to run code, test logic, and execute commands safely.

Planned

Improved Tools

Expanded tool integrations including secure local workspace file access and internet search APIs.

In Progress

Autonomous Agents

Multi-step agent loops executing tasks and self-correcting completely in the client tab.

In Progress

Offline Inference via Web Workers

Offload inference computation to Web Workers to ensure smooth UI responsiveness even under full load, supporting complete offline usage.

Developed locally with turborepo, Next.js, and Transformers.js.

All computing and model inference happen entirely client-side. Models are pulled directly from Hugging Face. No user queries or data is tracked or uploaded to any server.