GitHub – microsoft/NLWeb: Natural Language Web
As part of Microsoft Build 2025, Microsoft open sourced a new protocol called NLWeb, a collection of open protocols and associated open source tools
Imagine if every website could talk to you like ChatGPT. You type a question in plain English — “Do you sell gluten-free options?” or “When’s the next open day?” — and the site just answers.
That’s exactly what NLWeb by Microsoft enables.
But this isn’t “just another chatbot.” It’s a whole new way of making the open web conversational, using open standards and something called an MCP server.
Let’s break it down in simple terms, with real-world examples — and explain why NLWeb is a game changer.
What Is NLWeb?
NLWeb (short for Natural Language Web) is an open-source framework that lets websites become AI-friendly out of the box.
Instead of you having to build a custom chatbot, train it with your data, and host a service, NLWeb exposes your site’s content in a standard format so that any AI system can read and talk to it — safely and smartly.
At its core is a Model Context Protocol (MCP) server — a lightweight interface that lets AI agents understand and query your site’s content, similar to how APIs work, but made for AI.
💬 How Is NLWeb Different From a Regular Chatbot?
Let’s compare:
| 🛠️ Feature | Your Own Chatbot | NLWeb |
|---|---|---|
| Needs custom training? | ✅ Yes – you build & train it | ❌ No – reuses your existing structured data |
| Handles structured queries like AI agents? | ❌ Not usually | ✅ Yes – MCP allows it |
| Tightly bound to a specific AI? | ✅ Often (e.g. OpenAI plugin, RAG) | ❌ Works with any AI or agent |
| Uses open web standards? | ❌ Custom formats | ✅ Schema.org, RSS, JSON-LD, etc. |
| Automatically discoverable? | ❌ No | ✅ Yes – other systems can auto-query your MCP server |
So instead of building a chatbot about your site, NLWeb lets your site itself “speak AI.”
Real-World Example 1: A Local Bakery Website
Without NLWeb:
To let AI answer „Do you sell vegan cupcakes?“, you’d need to:
- Build a custom chatbot.
- Train it with product data.
- Host it somewhere.
With NLWeb:
You just expose your existing menu using Schema.org data, install an MCP server, and boom — your site can now respond to that question directly from the browser, AI assistant, or even a voice interface.
Bonus: NLWeb Makes Your Site AI-Discoverable
Here’s the fancy bit.
By setting up your MCP server (which NLWeb helps with), you’re not just helping your site talk to visitors — you’re enabling external AI agents to query your site too.
Imagine:
- A personal AI assistant asks, “When is my local recycling center open?”
- Instead of going to Google, it directly queries your city council’s website — because it has an NLWeb MCP server.
That means:
- Your site becomes part of the AI-driven web.
- You don’t need to manually integrate with every AI tool — they come to you.
🔐 Do I Have to Give Up Control?
Nope. You control what data is exposed and how it’s formatted. You can choose:
- What pages or datasets to make available.
- Which AI systems or bots can access it.
- Whether to rate-limit or secure access.
It’s your site, your rules — just smarter.
🔧 Getting Started (for Devs)
If you’re technical, here’s how you can explore:
- Add structured data to your site (Schema.org, JSON-LD, etc.).
- Install the NLWeb MCP server.
- Configure what content you want to expose.
- That’s it — you’re ready to be “AI-accessible.”
This is the flow from client to server:
Final Thoughts: Why NLWeb Matters
NLWeb is like HTML for the AI age. Just like HTML let browsers understand your website, NLWeb lets AI systems understand and talk to your website.
Whether you’re a bakery, a travel site, or a university — this is part of the future of search, support, and discovery.