2026-05-17 · KY · ~5 min read

Website MCP Endpoint: What It Is and How to Get One in 5 Minutes

You searched for "website MCP endpoint" because you want one. Here's what one is, what the URL actually looks like, and the fastest way to get one — about five minutes from clicking sign-in to pasting it into Claude or ChatGPT.

What a website MCP endpoint actually is

A website MCP endpoint is a single URL that lets an AI assistant — Claude Desktop, ChatGPT custom connectors, Cursor, Gemini CLI, Codex, anything that speaks the Model Context Protocol — query the live content of a website on demand. Paste the URL once into the AI client's connector settings; from then on the assistant can answer questions about that site using exact passages it pulls from a real index, with source URLs cited.

It is not a chatbot widget you embed on the website itself. It is the inverse — it lets your users' AI clients reach into your content from outside, on your terms. Same shape as letting Google index your pages, except the consumer is an LLM and the access is structured (streamable HTTP JSON-RPC) instead of HTML scraping.

The result, in one sentence: an AI asked about your site stops guessing and starts citing.

What the URL looks like

A WebToMCP endpoint URL is the form:

https://mcp.webtomcp.net/batches/<batch-id>/mcp?key=wmcp_<your-key>

Two parts matter. The batch-id is a UUID specific to one of your knowledge bases — a single WebToMCP account can hold several batches, one per logical knowledge base you keep separate (e.g. your public docs, your customer-only KB, and your changelog as three independent endpoints). The ?key= query parameter is a scoped access token; anyone with the full URL can query that endpoint, so treat it like a personal access token.

The endpoint speaks streamable HTTP MCP, which Claude Desktop, ChatGPT, and Cursor all support natively in 2026. For older clients that need stdio transport, the dashboard generates a paste-ready mcp-remote bridge command alongside the raw URL.

Get one in five minutes

Sign in. Open webtomcp.net/login and sign in with Google. Free tier, no credit card.
Add your website URL. Paste the URL of the site you want indexed — could be your docs (https://docs.example.com), your store (https://store.example.com), or your root domain. We respect robots.txt and only crawl within the path prefix you submit.
Wait for indexing. Most sites under 500 pages finish in 1–10 minutes. JavaScript-rendered pages are supported and metered per plan. The dashboard polls live status — you'll see pages indexed, chunks created, and vectors stored as the crawl progresses.
Copy the endpoint URL. When the batch flips to Ready, the dashboard shows the endpoint URL plus paste-ready connector snippets for Claude Desktop, ChatGPT, Cursor, Gemini CLI, and Codex.

What to do with the URL once you have it

The endpoint is your hand-off to any MCP client. The typical workflow per client:

Claude Desktop: Settings → Connectors → Add → paste the URL → restart Claude. Done. Full walkthrough →
ChatGPT (every tier including Free): Settings → Connectors → Add custom connector → paste the URL as the MCP server endpoint. Full walkthrough →
Cursor: Settings → MCP → add a server with the URL. Cursor picks it up automatically.
Anything else: If the client speaks MCP, the URL is enough. Setup snippets for Gemini CLI and Codex live in the docs.

Once connected, the AI can answer any question that a human reading the site cover to cover could answer — pricing tier differences, refund policy specifics, the API authentication flow, the latest changelog entries, onboarding steps in a given region. Every answer the AI returns carries source URLs from your indexed pages.

Is this the right move for your site?

A website MCP endpoint earns its keep when:

You publish content that people would otherwise have to read across many pages to get an answer (docs, changelogs, knowledge bases, wikis, product catalogs).
You'd rather an AI assistant cite your real pages than guess from stale training data or partial fetches.
Your audience already lives inside an AI client — they ask Claude or ChatGPT before they Google.

It's overkill for a five-page brochure site. It's clear-cut value for any docs site, e-commerce catalog, or knowledge base with more than a couple of dozen pages.

Frequently asked

Does it cost anything? Free plan covers 1 knowledge base, 1 website, ~500 AI questions per month, no credit card. Hobby at $9/month covers 5 KBs, 10 websites, 5,000 questions, and 100 JS-rendered pages. Full pricing breakdown →

Do I need to write code? No. Sign in, paste a URL, copy an endpoint URL. The MCP client is what consumes the endpoint — you don't host or maintain anything yourself.

Is the endpoint public? The URL embeds a scoped access token (?key=wmcp_…), so it acts like a personal access token — share with anyone you trust, don't post on a public page. The token is rate-limited at the edge. OAuth-gated mode is on the roadmap.

What happens when I update the site? Batches re-crawl weekly by default; you can also trigger a manual refresh from the dashboard for an instant re-index.

Is my content kept private? Indexed content stays on Cloudflare's edge, encrypted at rest, never shared with advertisers, never used to train models. Privacy policy →

Three live demo endpoints — Cloudflare Workers docs, Tailwind CSS docs, Hono docs — are available on the landing page. Copy any one into your AI client (no account needed) and ask it something. Once you've seen the citation-backed answer flow on someone else's docs, sign in and index your own URL.

web to MCP: what the phrase actually means — the brand explainer.
Connect WebToMCP to Claude Desktop — full step-by-step.
Connect WebToMCP to ChatGPT — works on every tier.
All setup guides — Cursor, Gemini CLI, Codex.

Questions? developer@webtomcp.net. Or sign in with Google to index your own site (free).