NLWeb

Natural Language Web · NLWeb protocol · Microsoft NLWeb

Definition

NLWeb is an open Microsoft project that adds a natural-language conversational interface to a website based on its Schema.org markup and RSS, and exposes the same interface to AI agents as an MCP server.

What NLWeb is

NLWeb is a set of open protocols and tools that lets you ask a website a question in natural language and get a structured answer. Microsoft positions it as the HTML of the agentic web: a common way for a site to talk both to people and to AI agents. It builds on data the site already has: Schema.org markup and RSS feeds. The authors stress that the code in the repository is an example implementation, not the only correct one.

How it works

The site loads its structured data — products, recipes, events, reviews — into a vector store.
A user or an agent asks a question, for example “lightweight two-person tent under 3 kg”.
NLWeb finds matching objects through semantic search, and an LLM composes the answer.
The answer arrives as JSON: a list of Schema.org objects with a name, a URL, a relevance score and a description.

Mode	What it returns
`list`	A list of the most relevant objects (the default mode)
`summarize`	The list plus a short summary
`generate`	An LLM answer to the question based on the objects found, as in classic RAG

The /ask endpoint serves the site’s conversational interface, and /mcp serves agents over MCP. The project works with a range of LLMs (OpenAI, Gemini, Anthropic, DeepSeek and others) and vector databases.

NLWeb and other ways to open a site to agents

Approach	What it gives the agent	Effort for the store
Schema.org	Structured data on the pages	Marking up templates
llms.txt	A Markdown map of key pages	One file
NLWeb	Natural-language questions answered from the catalogue, plus MCP	Deploying a service and an index
Your own MCP server	Any tools: search, basket, order	Development and maintenance

What it means for a retailer

Markup matters again. The accuracy of NLWeb answers depends directly on how complete your Product markup is: price, availability, brand, specifications.
Two channels from one base. The same index powers natural-language search for shoppers and MCP access for agents.
Data freshness. According to the project description, production deployments connect NLWeb to live databases rather than a copy of the content — otherwise prices and availability go stale.
Status. As of September 2026 NLWeb is an open project, not a mandatory standard; a sensible start is a pilot within your agent-readiness plan.

Frequently asked questions

Who created NLWeb?

The project was conceived and developed by R.V. Guha — the author of RSS, RDF and Schema.org — who joined Microsoft as a CVP and Technical Fellow. Microsoft introduced NLWeb at Build on 19 May 2025. The code is open under the MIT licence; as of September 2026 the GitHub repository sits in the nlweb-ai organisation, and the old microsoft/NLWeb address redirects there.

How is NLWeb related to MCP?

Every NLWeb instance is also an MCP server: besides the /ask endpoint it has /mcp, with the core MCP methods and a main tool called ask, to which an agent passes a natural-language question. The authors put it this way: NLWeb is to MCP and A2A what HTML is to HTTP. The project description says A2A support will follow.

Does NLWeb require new site markup?

No, that is the point: NLWeb relies on formats sites already publish — Schema.org and RSS, which the project's authors estimate are used by more than 100 million websites. The fuller and more accurate your Product markup — price, availability, specifications — the more accurate the answers. The data is loaded into a vector store; Qdrant, Milvus, Elasticsearch, Postgres, Azure AI Search and others are supported.

What is the status of NLWeb as of September 2026?

The project is open and active, but its authors describe their code as a demonstration implementation rather than the definitive solution. Cloudflare offers NLWeb deployment through AI Search as a public preview intended for experimentation. NLWeb has not become a mandatory standard for AI assistants: it is one way to open a site to agents, alongside a dedicated MCP server and llms.txt.