Show HN: A zero-knowledge, zero-overhead routing mesh using pure web primitives

An open-source project called bringYourOwnBrain released a web SDK that runs the Qwen2.5 0.5B open-weight model locally in the browser via WebGPU, with automatic WASM CPU fallback through Transformers.js, so developers can add AI NPCs to web games with a single script tag and no backend or API keys. The Apache 2.0-licensed SDK streams tokens directly into game logic, and the project claims zero server-side inference or telemetry, with no user prompts leaving the browser. The release also includes a physics-driven, interruptible local AI NPC demo and a vanilla JavaScript chat room demonstrating streaming token queues and cross-origin browser storage cache sharing.

bringYourOwnBrain is an open-weight local AI platform that makes running small open weight LLMs free and easy for both devs and users on WebGPU & WASM. No servers to deploy, no API keys to hide, and no token invoices. Compute belongs to the player. Add a single <script tag to your web game or app. Mount the connect button in any div or let it float bottom-right. Zero backend required. In this initial release, the player clicks "Connect Local AI". An isolated sandbox loads the lightweight Qwen2.5 0.5B more model options are being added regularly open model directly onto their graphics card via WebGPU with automatic WASM CPU fallback . Inference streams straight into your game logic at native frame rates. No user prompts ever leave the browser, and you never pay an AI API token bill. Experience what local open-weight intelligence feels like inside standard web games. To my knowledge, this may be the world's first physics-driven, interruptible local AI NPC s . Shove or get shoved mid-sentence to trigger an instant canned reaction or interruption. The game is a tip of the hat to a friend / dedication to Steven J. Horst and his 'old-ass' black cat, Midnight 2026 A minimalist vanilla JS chat room demonstrating streaming token queues, local connection states, and responsive typewriter rendering. For convenience with testing how the browser storage cache share operates in cross-origin manner , I have provided the demo chat on multiple domains: Tell us what you want to build and we can pre-build a prompt for you to make an AI powered browser game using your favorite large LLM e.g., Gemini, Claude, ChatGPT, etc . Just copy the mega-prompt into free Claude or Gemini, and generate your own physics-driven BYOB AI games. Tell us what AI game you think would be fun and get a free guide + instant one-shot prompt to make it and play it today. Zero accounts. Zero email. 100% Free. Drop this into any HTML page. Works in pure JavaScript, React, Canvas, Phaser, or Matter.js. php < -- 1. Load the BYOB Web SDK No extension needed -- <script src="https://bringyourownbrain.org/byob-web-sdk.js" </script < -- 2. Mount the connect button in your UI -- <div id="ai-connector" </div <script const brain = new BringYourOwnBrainWeb { container: document.getElementById 'ai-connector' , width: '220px', height: '44px', onState: state = console.log 'BYOB State:', state } ; brain.mount ; // 3. Stream or await responses directly from the player's GPU: async function talkToNPC playerText { brain.onChunk = token = renderSpeechBubble token ; const reply = await brain.chat { role: 'system', content: 'You are a sarcastic goblin bouncer. Answer in 1 short sentence.' }, { role: 'user', content: playerText } , { max tokens: 60, temperature: 0.7 } ; } </script Uses WebLLM to run open weights models on Apple Silicon, Nvidia, and AMD GPUs. If WebGPU is unavailable or disabled, the engine automatically falls back to CPU via Transformers.js with zero code changes. Built on radical local privacy and open-source transparency. bringYourOwnBrain.org operates zero server-side inference or telemetry pipelines. When a user interacts with an AI model: The BYOB Web SDK and connection bridge are open-source under the Apache 2.0 License .