Wiki · Concept · Last reviewed June 25, 2026

NLWeb

NLWeb, short for Natural Language Web, is Microsoft's open project for giving websites conversational interfaces and making selected site content addressable by AI agents through a standardized natural-language endpoint.

Category: Concept / Agent Infrastructure Updated: June 25, 2026 Tags: NLWeb, agentic web, MCP, website search, structured data

Definition

NLWeb is an open Natural Language Web project introduced by Microsoft on May 19, 2025. Microsoft describes it as a way for web publishers to create natural language interfaces for their own sites, using their chosen model and their own data. The practical aim is not to replace the web page, but to add an addressable question-answer layer beside the page, search box, feed, and API.

The project matters because it moves AI Search and Answer Engines closer to the publisher's own infrastructure. Instead of relying only on a centralized answer engine to crawl and summarize a site, NLWeb gives the site a way to answer questions directly from its selected content and, if enabled, expose that interface to agents.

Microsoft's announcement compares NLWeb to HTML in the emerging agentic web. That analogy should be treated as ambition, not settled fact. HTML is a mature document language with decades of browser, standards, accessibility, and security practice. NLWeb is newer protocol and implementation work for conversational retrieval and agent access.

Mechanism

NLWeb builds on data that many websites already publish, including Schema.org markup, RSS, JSONL, and related semi-structured formats. The project repository says NLWeb combines open protocols, Schema.org formats, and sample code so sites can implement conversational endpoints for human users and structured interaction for agents.

The NLWeb specification v0.55 defines a standardized interface for natural language interactions with agents. Its primary operation is ask, which lets a caller ask an NLWeb agent for information or request a task. It also defines await for asynchronous or long-running responses. The request object is organized around query, context, prefer, and meta; responses include a required _meta block and can return answer, elicitation, promise, or failure response types.

For HTTP transport, the specification maps NLWeb to POST /ask and POST /await. The repository's REST API documentation also describes implementation endpoints at /ask and /mcp, with /mcp returning answers in a form usable by MCP clients and supporting core MCP methods. The specification is transport-neutral: it says the same core JSON structures can move over HTTP, WebSockets, JSON-RPC, or agentic protocols such as Model Context Protocol.

Agent Context

For the Agent-Native Internet, NLWeb turns a website from a thing an agent scrapes into a thing an agent can ask. That changes the governance surface. A product catalog, event listing, recipe site, newsroom, help center, or civic archive can choose which structured data becomes queryable, what summaries are generated, what actions are exposed, and what context is preserved between turns.

Every NLWeb instance is described by Microsoft as also being an MCP server when a publisher chooses to expose it that way. That makes NLWeb adjacent to Agent2Agent Protocol, but not identical. NLWeb is about a website's conversational content endpoint; A2A is about peer agent collaboration; MCP is about tool and resource access.

Governance Use

A serious NLWeb deployment should document the content sources, ingestion schedule, structured-data vocabulary, retrieval index, model provider, response mode, citation behavior, logging, retention, and access controls. If it exposes actions, those actions need separate approval and authorization rules rather than being bundled into search.

NLWeb also creates publisher-side evidence. Logs can show which query was asked, what data was eligible for retrieval, what result items were returned, what summary was generated, and whether an agent or human client made the call. Those records help resolve disputes about stale answers, missing sources, hallucinated summaries, and agent-initiated use.

Limits

NLWeb is not proof that a site is trustworthy, current, licensed for AI reuse, or safe for agent action. The specification says authentication and authorization are handled at the transport layer and are outside its scope. Rate limiting is also a transport-layer concern. A site that exposes NLWeb still needs ordinary security controls, permissioning, abuse monitoring, privacy review, and rollback.

The pattern can also amplify bad site data. If a catalog, feed, schema object, or support article is stale, promotional, or incomplete, the conversational surface may make that weakness sound more coherent. NLWeb makes website content more askable; it does not make the content true.

Review Record

Source Discipline

Claims about the announcement, project purpose, MCP framing, and early collaborators should cite Microsoft. Claims about protocol objects, response types, transport neutrality, authentication boundaries, rate limiting, and HTTP mapping should cite the NLWeb specification. Claims about implementation endpoints, vector stores, LLM connectors, and repository modules should cite the GitHub repository or its documentation.

Spiralist Reading

Spiralism reads NLWeb as the site learning to speak back to the agent. That can protect publishers from being flattened into someone else's answer box, but only if the speech remains accountable. A website should not become a disembodied oracle. It should answer with sources, limits, freshness, and a clear path back to the underlying record.

Sources

Return to Wiki