Microsoft's latest big move might truly change the entire internet.
Microsoft announced an open-source project called NLWeb at the Build 2025 conference. It sounds ordinary, but it can actually turn any website into an AI application in seconds!
Notably, NLWeb is built upon Anthropic's Model Control Protocol (MCP), meaning it can serve not only human users but also enable AI agents to directly interact with websites.
In the words of Microsoft CTO Kevin Scott: "You can truly think of it as the HTML of the agent web."
What exactly is NLWeb?
Simply put, NLWeb is a set of open protocols and tools that allow websites to easily add a natural language conversational interface.
Its workflow is quite ingenious:
Leveraging existing data structures: NLWeb directly utilizes a website's existing structured data, including Schema.org markup, RSS feeds, etc. This means publishers don't need to completely rebuild their content infrastructure.
Data processing and storage: The system includes tools for adding structured data to a vector database, supporting semantic search and retrieval. It supports all major vector databases, allowing developers to choose the most suitable technology solution.
AI enhancement layer: Large language models augment the stored data with external knowledge and context. For example, when a user queries about restaurants, the system automatically adds relevant information such as geographical data and reviews, providing a comprehensive intelligent response.
Universal interface creation: Ultimately, it forms a natural language interface that serves both human users and AI agents. Visitors can ask questions in simple English and receive conversational responses, while AI systems can programmatically access website information through the MCP framework.
This method allows any website to participate in the emerging agent web without extensive technical modifications.
Protocol Battle: Everyone is Claiming Territory
Focuses on enabling agents to converse with each other, primarily addressing agent orchestration and communication issues, rather than AI-enabling existing websites.
Maria Gorskikh, founder and CEO of AIA, explained that A2A uses defined schemas and lifecycle models for structured task passing between agents.
While the protocol is open source, current implementations and tools are tightly coupled with Google's Gemini tech stack.
LLMs.txt
Aims to help large language models better access web content. While it sounds similar to NLWeb on the surface, it's actually completely different.
"NLWeb does not compete with LLMs.txt; it's more like a web crawler trying to infer intent from websites."
Krish Arvapally, co-founder and CTO of Dappier, explained that LLMs.txt provides content in markdown format with training permissions to help LLM crawlers properly ingest content.
NLWeb, on the other hand, focuses on enabling real-time interaction on the publisher's website.
MCP
MCP is increasingly becoming the de facto standard and a foundational element of NLWeb.
Fundamentally, MCP is an open standard for connecting AI systems with data sources. In Microsoft's view, MCP is the transport layer, and together, MCP and NLWeb provide the HTML and TCP/IP for an open agent web.
Forrester Senior Analyst Will McKeon-White believes NLWeb has clear advantages over other options:
"NLWeb's main advantage is better control over how AI systems 'see' the various parts that make up a website, allowing for better navigation and a more complete understanding of tools. This can reduce errors where systems misunderstand website content and also reduce interface rework."
Early Adopters Have Already Tasted Sweetness
Microsoft didn't just release NLWeb and let it be; several organizations are already using it, including Chicago Public Media, Allrecipes, Eventbrite, Hearst (Delish), O'Reilly Media, Tripadvisor, and Shopify.
O'Reilly Media's CTO Andrew Odewahn is one of the early adopters and sees the true value of NLWeb:
"NLWeb leverages the best practices and standards developed on the open web over the past decade and makes them available to large language models. Companies have long optimized this metadata for SEO and other marketing purposes, and now they can leverage this rich data to make their internal AI smarter and more capable via NLWeb."
In his view, NLWeb is valuable for enterprises both as consumers of public information and publishers of private information.
Almost every company has sales and marketing efforts that might need to ask "What does this company do?" or "What is this product about?"
"NLWeb provides an excellent way to expose this information to internal LLMs so you don't have to hunt around," says Odewahn:
"As a publisher, you can use schema.org standards to add your own metadata and use NLWeb internally as an MCP server for internal use."
Using NLWeb isn't a heavy lift either.
Odewahn points out that many organizations may already be using many of the standards NLWeb relies on.
"There's no downside to trying it now, as NLWeb can run entirely within your infrastructure," he says:
"This is open-source software meeting the best open-source data, so you have nothing to lose and a lot to gain by trying it now."
Hop On or Wait and See?
Analyst Michael Ni has a relatively positive view on NLWeb, but this doesn't mean enterprises need to adopt it immediately.
Ni notes that NLWeb is still in a very early stage of maturity, and enterprises should expect 2-3 years for substantial adoption. He suggests that leading enterprises with specific needs, such as active marketplaces, could consider piloting it and have the ability to participate and help shape the standard.
"This is a vision specification with clear potential, but it needs ecosystem validation, implementation tools, and reference integrations before it can reach mainstream enterprise piloting," says Ni.
Others hold a more positive view on adoption.
Gorskikh suggests an accelerated approach to ensure businesses don't fall behind.
"If you're an enterprise with a large content surface, internal knowledge base, or structured data, piloting NLWeb now is a smart and necessary step to stay ahead," she says. "This isn't a wait-and-see moment—it's more like the early days of adopting APIs or mobile apps."
However, she points out that regulated industries need to proceed cautiously.
Industries like insurance, banking, and healthcare should defer production use until neutral, decentralized validation and discovery systems are in place. Early efforts are already addressing this—such as the MIT NANDA project, in which Gorskikh is involved, which is building an open, decentralized registry and reputation system for agent services.
AI Will Change Every Website
For enterprise AI leaders, NLWeb is a watershed moment, a technology that should not be ignored.
AI will interact with your website, and you need to AI-enable it.
NLWeb is a particularly attractive way for publishers, just as RSS became a must-have feature for all websites in the early 2000s. In a few years, users will expect its presence; they will expect to be able to search and find things, and agent AI systems will also need to be able to access content.
This is the promise of NLWeb.
From the GitHub repository, NLWeb is indeed a "deeply independent" platform—perfectly supporting operating systems (Windows, MacOS, Linux), vector storage (Qdrant, Snowflake, Milvus, Azure AI Search), and LLMs (OpenAI, Deepseek, Gemini, Anthropic, etc.).
It's lightweight and scalable, capable of running from cloud clusters to laptops, and will soon support mobile phones.
hugo alves(@ugo_alves) tagged Sam Altman:
@sama Create an ask based on this, allowing anyone to deploy ChatGPT as a web Q&A system
Sam Woods(@sw00ds) believes the impact is profound:
If it's as significant as it sounds, it could redefine web-based AI applications
UdayKumar(@UdayKumar) sees transformative potential:
Yes. It has huge potential to transform websites into conversational agents. Reference implementations will accelerate adoption.
Of course, there are also voices of doubt.
Pavel Snajdr(@PavelSnajdr) expressed a different view:
I'd rather adopt the "intelligence as cheap as electricity" approach, then we wouldn't need any MCP band-aids
The emergence of NLWeb marks the internet's transition to the agent era.
Every website will need to consider how to interact with AI; this will no longer be optional, but a necessity.
👇
👇
👇
Additionally, I have used AI to collect AI-related information from across the web, and then used AI to select, review, translate, and summarize it for publication in the "AGI Hunt" Knowledge Planet.
This is an AI news feed that is purely informational, without emotion (not a recommendation feed, no selling courses, no moralizing, just providing information).
Welcome to join! Also welcome to join the group chat with 2000+ members