home / comparison
WebMCP vs MCP vs llms.txt: the agentic web stack, untangled
Four standards keep getting confused: MCP, WebMCP, llms.txt, and NLWeb. Here's what each layer actually does, who's behind it, and what your website should implement first — in one comparison matrix.
Updated 2026-08-28 · webmcp-ai.dev research desk
The "agentic web" conversation is drowning in acronyms. MCP, WebMCP, llms.txt, NLWeb, plus Schema.org and MCP server cards — each covers a different layer, none replaces the others, and most advice online mixes them up. Here is the clean mental model: agents need to find you, understand you, and act on you. Each standard serves one of those.
The comparison matrix
| Standard | Layer | Who's behind it | Where it lives | What it gives agents |
|---|---|---|---|---|
| MCP Model Context Protocol |
Backend capability | Anthropic (2024), now industry-wide | A server you run, out-of-band from your site | Tools/resources for AI assistants (Claude, IDEs, agents) with their own auth |
| WebMCP | In-page capability | Google + Microsoft → W3C WebML CG (2025) | Inside your web page, via document.modelContext |
Callable tools in the user's own browser session — no API keys, human in the loop |
| llms.txt | Identity & content discovery | Community convention (Jeremy Howard, 2024) | /llms.txt at your site root |
A curated markdown map of what your site is and where key content lives |
| NLWeb | Conversational interface | Microsoft (2025) | A protocol/endpoint on your site | Natural-language query interface over your content, MCP-compatible |
| Schema.org JSON-LD | Comprehension | Google/Microsoft/Yahoo/Yandex (2011) | Structured data in your HTML | Machine-readable meaning: products, prices, hours, FAQs, reviews |
| MCP server card | Capability discovery | MCP ecosystem convention | /.well-known/mcp/server-card.json |
Tells agents where your backend MCP server is and what it offers |
MCP vs WebMCP: the distinction that matters
Both let models call tools, which is why they get confused. The difference is where the tools run and whose session they use:
- An MCP server is backend infrastructure. You build and host it, design its auth (API keys, OAuth), and assistants connect to it directly — no browser involved. Great for B2B integrations and data access at scale.
- A WebMCP tool is three lines of frontend JavaScript wrapping functionality your site already has. It executes in the visitor's browser, as the logged-in visitor, with the browser mediating consent. Great for consumer actions: search, book, add to cart, check out.
Rule of thumb: MCP is for your API; WebMCP is for your UI. A mature agent strategy eventually ships both, but for most websites WebMCP is the faster win because it reuses the frontend you already have.
What should you implement first?
- Foundation (hours): server-rendered content, semantic HTML, Schema.org JSON-LD, robots.txt rules that deliberately allow the AI crawlers you want.
- Identity (an afternoon): llms.txt describing your site's purpose and key pages.
- Capability (days): WebMCP tools for your top one or two conversion actions — the ones an agent would perform on a user's behalf. See tool patterns by industry.
- Scale (when justified): a backend MCP server and server card, if partners or assistants need out-of-band access to your data.
Our deep audit checks all of these layers on your actual site and sequences them into a roadmap — or start with the free scan.