# WebML

> WebML (webml.ai) is a browser AI lab by Omri Beladev. It runs machine learning models entirely in the browser — local inference, client-side RAG, WebGPU compute, on-device embeddings, and a three-worker voice agent. No server, no API key, no cloud.

Key facts:
- What is WebML? WebML (webml.ai) is a browser AI lab that runs machine learning models entirely in the browser using WebGPU, ONNX Runtime Web, and Transformers.js — with no server, no API key, and no cloud required. It includes live experiments, articles, and an interactive playground.
- What is browser-native AI? Browser-native AI runs inference directly inside a web browser using WebGPU compute shaders and WebAssembly. Models execute on the user's device GPU or CPU, with no data leaving the browser — enabling privacy-first, offline-capable AI applications.
- What technologies does WebML use? WebML experiments use Transformers.js (Hugging Face), WebLLM (MLC-AI), ONNX Runtime Web, Chrome's built-in Prompt API (Gemini Nano), Web Workers for threading, AudioWorklet for audio processing, and WebGPU for GPU-accelerated inference.
- Does WebML work offline? Yes. Once model weights are downloaded and cached via the Cache API or Origin Private File System (OPFS), all inference runs locally with no network connection needed.

## Pages

- [Home](https://webml.ai/): Overview of the lab, live runtime detection, experiments, and articles.
- [Playground](https://webml.ai/playground): Chat with an LLM running 100% in your browser via WebLLM and WebGPU.
- [Lab](https://webml.ai/lab): All browser AI experiments.
- [Articles](https://webml.ai/articles): Technical writing on browser-native AI.
- [About](https://webml.ai/about): About Omri Beladev, the engineer behind WebML.

## Experiments

- [LLM Playground](/playground) (live): Chat with Phi-3.5, Llama 3.2, Gemma 2, or Mistral — running 100% in your browser via WebGPU. No API key, no server, nothing leaves your tab.
- [Todo AI](https://todo-ai-taupe.vercel.app/) (alpha): A todo list with an embedded AI chatbot running entirely in the browser. No server, no API key — compare WebLLM and Chrome's Prompt API side by side.
- [Detectiks](https://detectiks.com) (live): Detect whether an image was generated by AI or captured by a human. Real-time, in the browser.
- [webmcp.js](https://webmcp.webml.co.il/) (alpha): A browser-side library that turns any website into a structured capability layer for AI agents.

## Articles

- [LiteRT.js Moves the TFLite Runtime Into the Browser](https://webml.ai/articles/litert-js-browser-runtime): Google released LiteRT.js as a web binding for its LiteRT stack, giving developers a direct path from .tflite models to browser inference. What changes for TensorFlow.js apps, PyTorch conversion, hardware acceleration, and model choice. (Published 2026-07-10)
- [WebMCP's Second API Skips JavaScript Entirely. It's Two HTML Attributes.](https://webml.ai/articles/webmcp-declarative-forms): Chrome's WebMCP origin trial (sign-up opened in Chrome 149, June 9) ships a Declarative API alongside the Imperative one: annotate a form with toolname, tooldescription, and toolparamdescription and the browser builds the tool from markup you already have. What toolautosubmit, respondWith(), and the tool-form-active CSS hooks give you, and where it stops working. (Published 2026-07-01)
- [A Browser Voice Agent Built on Three Workers](https://webml.ai/articles/browser-voice-agent-three-workers): How the WebML voice agent wires three on-device AI models together in the browser using Web Workers, AudioWorklet, and constrained JSON generation. The architecture decisions behind running real-time speech-to-action locally. (Published 2026-06-06)
- [Chrome's Prompt API Now Runs on Web Pages. The Spec Process Disagrees.](https://webml.ai/articles/prompt-api-web-pages-chrome-148): What changed in Chrome 148 for the Prompt API: the LanguageModel global is now available to web pages, with multimodal input, structured output constraints, and real hardware requirements — plus why the browser standards community formally opposed it. (Published 2026-06-06)
- [Local LLM in a Todo App — WebLLM vs Prompt API](https://webml.ai/articles/todo-ai-local-llm): A hands-on comparison of WebLLM and Chrome's Prompt API, built into a practical Angular todo app. No server, no API key — the model runs in the browser. (Published 2026-05-25)
- [WebMCP Needs a Browser Tool Layer — So I’m Building webmcp.js](https://webml.ai/articles/webmcp-browser-tool-layer): An introduction to webmcp.js, a small browser-side library for exposing website capabilities as structured, approval-aware tools for AI agents. (Published 2026-05-24)
- [The runtime behind browser AI](https://webml.ai/articles/browser-ai-runtime): A technical walkthrough of the browser AI execution stack: ONNX Runtime Web, the WASM and WebGPU backends, model caching via Cache API and OPFS, Web Worker threading, and how WebLLM differs from Transformers.js under the hood. (Published 2025-05-20)

## Optional

- [Full article text](https://webml.ai/llms-full.txt): Complete plain-text bodies of all articles.
- [RSS feed](https://webml.ai/feed.xml)
- Author: Omri Beladev — GitHub: https://github.com/beladevo, LinkedIn: https://www.linkedin.com/in/omribeladev/