NLWeb

NLWeb is a set of open protocols and reference tools for adding natural-language interfaces to websites, usable by both people and AI agents. It builds on Schema.org and related semi-structured formats such as RSS, which many sites already publish, and treats that markup as a semantic layer that a conversational endpoint can query.

It has two parts. The first is a protocol for asking a site questions in natural language, with responses returned as JSON using Schema.org vocabulary. The second is a reference implementation that works over sites with structured lists (products, recipes, attractions, reviews) and pairs with UI widgets for a conversational front end. Every NLWeb instance is also an MCP server exposing a core ask method, with A2A support planned. The project describes its relationship to MCP and A2A as like HTML's to HTTP.

The code is offered as proof-of-concept implementations rather than a definitive solution. The repository contains AskAgent (the core query agent, with LLM and vector-store connectors, data ingestion tools, and a sample web UI), AgentFinder (discovery and routing across NLWeb agents), DataFinder (natural language to SQL for sources such as HubSpot, Dynamics 365, and Jira), ModelRouter (selects cost-effective models that meet quality thresholds), and NLWebScorer (neural models for ranking result quality).

It runs on Windows, macOS, and Linux, from laptops to data-center clusters. Supported vector stores are Qdrant, Snowflake, Milvus, Azure AI Search, Elasticsearch, Postgres, and Cloudflare AutoRAG, and supported LLM providers are OpenAI, DeepSeek, Gemini, Anthropic, Inception, and Hugging Face. Prompts, control flow, the UI, and memory are configurable, and Azure is a documented deployment target.