Skip to content
Case study · Model Context Protocol

A public MCP server AI assistants can call

People increasingly ask an AI assistant about a supplier before they visit its website. We built a Model Context Protocol server so Claude, ChatGPT and IDE assistants can read current, citable answers about Radiatus, and use a few developer utilities, straight from the source.

Abstract 3D letters A and I representing AI assistants using tools
Case study at a glance
  • 10 toolsAll read-only
  • 4 protocol versions2026-07-28 and three 2025 releases
  • 120 per minuteRequests per client IP
  • No IPs loggedHashed for rate limiting only

The problem

AI assistants answer questions about companies from whatever they scraped, often months old. Our service list, technology pages and guides change every few weeks, and we wanted assistants to quote the current version with a link, not a guess.

What we built

A stateless MCP endpoint at radiatus.com/mcp with ten read-only tools: three that search and return Radiatus pages, one WhatsApp cost estimator for India, and six deterministic developer utilities.

Where it stands

Public and free, with no sign-up. It answers from the same catalogue and page files that build the website, so an assistant and a visitor see the same facts on the same day.

Numbers you can check

Every figure on this page comes from the running product or its source code, checked on 7 November 2026. None is a client result.

FigureWhat it countsWhere it comes from
10Read-only tools: search, page lookup, offerings list, WhatsApp cost estimate, hashing, encoding, JWT decoding, UUIDs, timestamps, IPv4 subnetsThe server's tool registry, each marked read-only
4MCP protocol versions answered: 2026-07-28 plus 2025-11-25, 2025-06-18 and 2025-03-26 for older clientsThe server's version lists
120Requests per minute allowed per client, keyed on a hash of the IP addressThe rate limiter
64 KBLargest request body acceptedThe request parser
0Sessions stored: every request is served statelesslyThe transport implementation

Why a company website needs an MCP server

The Model Context Protocol lets an AI assistant call tools on a remote server and use the results in its answer. For a services company, that turns "what does Radiatus do for HL7 integration?" from a guess based on an old crawl into a lookup of the current page, returned with its URL so the assistant can cite it.

We also run llms.txt and a full-text version for AI crawlers. The MCP server goes a step further: an assistant can search, fetch a specific page or list our offerings by category on demand.

Design decisions

Stateless Streamable HTTP

The server accepts POST requests and answers each with a single JSON object. Every tool returns immediately, so there is no need for server-sent events, and no session is created, which keeps it simple to run behind an ordinary PHP stack and impossible to exhaust with abandoned sessions. GET and DELETE return 405, and JSON-RPC batches are rejected.

Old and new clients at once

MCP clients in the wild speak different protocol versions. Modern clients using the 2026-07-28 revision send their version and method in headers, which the server checks against the request body, and can call server/discover. Older clients still send initialize; the server negotiates one of three 2025 versions and then serves them the same tools statelessly.

One source of truth

Tool answers are built from the same catalogue and page files that render radiatus.com. When a page is updated, the MCP answer changes with it. There is no second copy of the content to drift out of date.

Safe to leave open

A public endpoint with no authentication is only safe if it cannot do harm. Every tool is read-only and side-effect free. Requests are limited to 120 per minute per client, the body is capped at 64 KB, malformed Origin headers are refused, and the endpoint tells search engines not to index it. For usage statistics the server logs the tool name, protocol era and client name. IP addresses are hashed for the rate limiter and never written to the log.

The stack

LayerChoiceWhy
TransportMCP Streamable HTTP, single JSON responsesNo streaming needed when every tool answers instantly
RuntimePHP 8.3 behind Apache, alongside the websiteReuses the site's catalogue and copy files
Protocol2026-07-28 with server/discover; 2025-11-25, 2025-06-18, 2025-03-26 via initializeWorks with current and older assistants
ProtectionPer-IP rate limit (hashed), 64 KB body cap, Origin check, noindexPublic without being abusable

What it does not do yet

  • Tools only read: the server cannot book a call, send a message or change anything.
  • No MCP resources or prompts are published, only tools.
  • Responses are not streamed; long-running tools would need a different transport design.

What this means for your own MCP server

Most organisations need the opposite of a public server: a private one that exposes internal APIs and data to their own assistants, with OAuth, per-user permissions and audit logs. The decisions above (stateless where possible, read-only by default, one source of truth, careful logging) carry over. Our MCP server development page explains how we build them, and the generative AI and agents practice covers the assistants that use them.

FAQ

Questions about this case study

Add https://radiatus.com/mcp as a remote MCP server (Streamable HTTP) in an MCP-capable client such as Claude, ChatGPT, VS Code or Cursor. No sign-up or API key is needed.

The 2026-07-28 revision for modern clients, and 2025-11-25, 2025-06-18 and 2025-03-26 for clients that still use the initialize handshake.

All ten tools are read-only and have no side effects. Requests are rate-limited, and the server does not log IP addresses.

Yes. We build MCP servers that expose your APIs and data to AI assistants with OAuth, scoped permissions and audit logs, and we test them against the clients your team uses.

Work with the team that built it

Need something like this for your organisation?

Tell us what you are connecting, running or securing. The engineers behind this case study reply personally.

  • A senior engineer reads every enquiry
  • Happy to sign an NDA first
  • You get a written next step, not a sales sequence
Last updated