What is WebLLM?

WebLLM is building the future of AI on the web - where users control their own AI, developers build without backend costs, and AI becomes a standard web capability like graphics, video, or authentication.

The Core Idea: Bring Your Own AI (BYOAI)

What if AI worked like printing?

When you print from a website:

• The website doesn’t need to know about your printer
• You configure your printer once in your OS
• Every website can print through the same interface
• Websites don’t pay for your printing

WebLLM makes AI work the same way:

• Websites don’t manage AI providers
• You configure your AI once in your browser
• Every website uses AI through the same interface
• Websites don’t pay for your AI usage

This is Bring Your Own AI (BYOAI) - users provide the AI, websites provide the features.

Why This Matters

The Current Broken Model

Today, when websites want AI features:

User → Website Server → AI Provider
      ↑                ↑
   Website pays     Website's API key

Problems:

1. Cost barrier - Websites pay for every user’s AI usage ($1000s/month)
2. Privacy loss - Your data flows through website servers
3. Limited access - Only big companies can afford AI features
4. Vendor lock-in - Websites pick the provider, you’re stuck with it
5. Redundant infrastructure - Every website rebuilds the same thing

The WebLLM Model

With WebLLM:

User → WebLLM → [User's Choice:]
               ├─ Free local model
               ├─ Claude (user's API key)
               └─ GPT-4 (user's API key)
    ↑
Website never touches your data or pays for usage

Benefits:

1. Zero costs for websites - Users provide their own AI
2. Maximum privacy - Data goes where you choose (or nowhere)
3. Universal access - Any website can add AI features
4. User choice - Use any model, any provider
5. One implementation - Websites use standard API

Inspiration from Web Standards

WebLLM follows the pattern of successful web standards:

WebGL: Standardized Graphics

Before: Flash, plugins, proprietary solutions After: One API, works everywhere, hardware-accelerated Impact: 3D graphics became accessible to all websites

WebRTC: Standardized Video/Audio

Before: Skype plugin, proprietary solutions After: Video calls work in any browser Impact: Video conferencing became a web platform feature

WebAuthn: Standardized Authentication

Before: Passwords everywhere, different 2FA systems After: Biometric auth, security keys, unified API Impact: Stronger security, better UX

WebLLM: Standardized AI

Before: Every website builds AI backend After: One API, user-controlled providers Impact: AI becomes accessible to all websites

How It Works Today

We’re starting with a browser extension that proves the concept:

For Users

1. Install extension - Add WebLLM to Chrome
2. Configure providers - Choose local models, add API keys, or both
3. Use AI everywhere - Visit WebLLM-enabled websites

Your configuration works across all websites - set it once, use everywhere.

For Developers

1. Install client library - npm install webllm
2. Call the API - Simple, provider-agnostic code
3. Ship features - Zero infrastructure, zero costs

import { WebLLMClient } from 'webllm';

const client = new WebLLMClient();
const summary = await client.generate({
  prompt: 'Summarize: ' + articleText
});

That’s it. No servers, no API keys, no billing.

For the Ecosystem

Websites that add WebLLM:

• Get AI features without infrastructure
• Attract users who want to use their own AI
• Reduce costs dramatically
• Improve privacy by design

The Path to Standardization

Our journey to making this a web standard:

Today: Extension + SDK

Status: Available now

• Chrome extension managing providers
• Developer SDK for websites
• Real-world testing and feedback
• Community building

Near-Term: Prove It Works

Goal: Demonstrate the value

• Thousands of websites using WebLLM
• Hundreds of thousands of users
• Clear benefits in privacy and cost
• Refined API based on real usage

Medium-Term: Formal Proposal

Goal: Submit to W3C

• Complete technical specification
• Security and privacy analysis
• Browser vendor discussions
• Origin trials in browsers

Long-Term: Native Integration

Goal: Built into browsers

• No extension needed
• Faster, more integrated
• Universal standard
• Like WebGL, WebRTC, WebAuthn

See the full timeline: WebLLM Roadmap

Key Principles

1. User Control First

Users always decide:

• Which AI provider to use (or multiple)
• Where their data goes (local, cloud, or hybrid)
• How much to spend (free local, pay-per-use cloud)
• What permissions to grant

Websites never control this.

2. Privacy by Default

Privacy is built-in:

• Local models process data on your device
• Cloud providers get data directly from extension (not website)
• Transparent data flows
• No tracking or telemetry by WebLLM

Your data, your choice.

3. Open Standard

No lock-in:

• Open specification (W3C track)
• Open source implementation
• Works across browsers (eventually)
• Community-driven development

Not controlled by any company.

4. Developer Experience

Make it simple:

• 3 lines of code to add AI
• Provider-agnostic API
• Great documentation
• Free to use

Remove barriers to innovation.

Real-World Benefits

For Privacy-Conscious Users

Use case: You want AI features but don’t trust cloud services

Solution:

• Download a local model (free, runs on your computer)
• All processing happens offline
• Data never leaves your device
• Works without internet

For Power Users

Use case: You already pay for Claude Pro or ChatGPT Plus

Solution:

• Use your API keys across all WebLLM websites
• One subscription, unlimited sites
• Premium models everywhere
• Cost-effective

For Indie Developers

Use case: You want AI features but can’t afford $1000+/month

Solution:

• Users bring their own AI
• You pay $0 for AI infrastructure
• Ship features competitors can’t afford
• Compete with big companies

For Enterprises

Use case: You need AI but have strict compliance requirements

Solution:

• Use local models for sensitive data
• Full audit trail of all AI usage
• No PII sent to third parties
• Simplified GDPR/CCPA compliance

For Students & Educators

Use case: Need AI for learning but budgets are tight

Solution:

• Free local models for everyone
• No per-student costs for schools
• Privacy-respecting by design
• Accessible to all

What Makes WebLLM Different?

vs. Traditional AI Integration

Traditional:

• Website builds backend
• Website pays for API
• Website chooses model
• Your data flows through their servers

WebLLM:

• No backend needed
• User pays (or uses free local models)
• User chooses model
• Data never touches website servers

vs. Chrome Built-in AI

Chrome AI:

• One weak model
• No choice
• Chrome only
• Not extensible

WebLLM:

• Any model (local or cloud)
• User chooses
• Works across browsers (goal)
• Open standard

vs. Direct API Usage

Direct API:

• Every website implements separately
• Redundant infrastructure
• Different APIs to learn
• No user control

WebLLM:

• One standard API
• Shared infrastructure (extension)
• Same API everywhere
• User control by design

The Vision: AI as a Web Platform Feature

In 5 years, we envision:

For Users:

• AI settings in browser (like camera/mic permissions)
• Download models once, use everywhere
• Configure providers in browser settings
• Transparent, controlled AI usage

For Developers:

• navigator.llm works in all browsers
• No extension needed
• Standard web development curriculum
• Default choice for AI features

For the Web:

• AI is a standard capability
• Open, interoperable, universal
• Privacy-respecting by design
• Accessible to everyone

Getting Started

Ready to experience AI on your terms?

As a User

➡️ Install Extension - Get started in 5 minutes ➡️ Privacy Benefits - Learn how your data is protected

As a Developer

➡️ Developer Guide - Add AI to your website ➡️ API Reference - Technical documentation

Learn More

➡️ Roadmap - Where we’re headed ➡️ Architecture - How it works technically ➡️ Extension Details - Current implementation

Questions

Is this production-ready?

Yes! The extension is stable and ready to use. Websites are using it in production. The standardization journey is just beginning, but the value is available today.

Will this really become a standard?

We’re following the proven path. It takes years, but we’re optimistic. Meanwhile, the extension delivers value now.

What if I don’t want to manage AI providers?

WebLLM works great with defaults:

• Just download one local model (free, easy)
• Or use a single API key (set once, forget)
• The extension handles everything else

Can websites abuse this?

No. You always grant permission first. You can revoke anytime. You can set limits. Full audit trail.

What’s the catch?

No catch. This is how the web should work - user-controlled, privacy-respecting, open standards.

---

AI should work for users, not just companies. WebLLM makes that real.