NLWeb
NLWeb is an open-source project from Microsoft, announced at Build in May 2025, that turns a website's existing structured data — schema.org markup, RSS feeds, JSON feeds — into a natural-language interface an AI agent can query directly. It does not require a new feed format; it reuses what a site already publishes. Every NLWeb instance also runs as a Model Context Protocol (MCP) server, so any MCP-compatible agent (Claude, ChatGPT, and others) can call it without a bespoke integration.
What NLWeb actually does
NLWeb is a Python project Microsoft open-sourced and announced at Build 2025. Microsoft CTO Kevin Scott described the goal as making it easy for "anyone who has a website or an API already" to turn it into an agentic application — "a little bit like HTML for the agentic web."
Mechanically, it sits on top of data a site is usually already producing: product schema markup, RSS, JSON feeds, sitemaps. NLWeb indexes that content and adds a conversational query layer over it, so a visitor — human or AI agent — can ask a question in plain language ("which of these have a 1/2-inch NPT port") instead of clicking through facets. Microsoft's launch partners included Shopify, Tripadvisor, and Eventbrite; Cloudflare added native, managed NLWeb hosting through its AutoRAG infrastructure in early 2026, which lowered the bar for sites that don't want to run the Python service themselves.
How it relates to MCP and schema markup
Three layers get conflated here, and they are not the same thing:
| Layer | What it is |
|---|---|
| Schema.org / JSON-LD | The structured data a page carries about itself |
| MCP | The wire protocol an agent uses to call a tool or read a resource |
| NLWeb | An application built on both — it reads a site's structured data and exposes a query interface over MCP |
NLWeb is not a replacement for product schema markup; it depends on it as input. It is also not a competitor to llms.txt so much as a different tier of the same idea — llms.txt is a static pointer file an AI crawler reads once; NLWeb is a live, queryable interface that answers a specific question at request time.
What a catalog needs before NLWeb is worth turning on
NLWeb answers with whatever the underlying schema markup and feeds already contain. It does not add a missing attribute, infer an application use case, or translate supplier nomenclature into buyer language — it retrieves and summarizes what is already structured. A catalog with thin or inconsistent Product schema gets an NLWeb interface that returns thin, inconsistent answers, just faster and in a chat window instead of a search results page.
For a distributor evaluating NLWeb, the honest prerequisite is the same one every AI-search protocol has: complete, accurate, buyer-language product attributes published in a machine-readable form. Anglera's enrichment work sits upstream of this — it is the step that makes the schema markup NLWeb (or any agent) reads worth querying in the first place.
Frequently asked questions
What is NLWeb?
NLWeb is an open-source project Microsoft announced at Build 2025 that lets a website expose its existing structured data — schema.org markup, RSS, JSON feeds — as a natural-language interface. Every NLWeb instance runs as a Model Context Protocol (MCP) server, so AI agents can query it directly.
Do I need a new feed format to support NLWeb?
No. NLWeb is built to reuse structured data a site is usually already producing — schema.org markup on product pages, RSS, and JSON feeds — rather than requiring a separate NLWeb-specific format.
Is NLWeb the same as MCP?
No. MCP is the general-purpose wire protocol that lets an AI agent call a tool or read a resource. NLWeb is an application built on top of MCP and on a site's existing structured data — it uses MCP as the mechanism agents use to reach it, but it is not the protocol itself.
Who has adopted NLWeb so far?
Microsoft named Shopify, Tripadvisor, and Eventbrite as early partners at launch. Cloudflare added managed, native NLWeb hosting through its AutoRAG infrastructure in early 2026, giving smaller sites a path that doesn't require self-hosting the underlying Python service.