MCP for business websites: a plain-English guide
The Model Context Protocol is how AI assistants plug into outside services. Give your business its own MCP server and an assistant can ask you questions, send you an inquiry or book a meeting directly, instead of fumbling through your website. Here’s what that means, what it takes, and who can use it today.
Updated
What MCP is, in plain English
The Model Context Protocol (MCP) is an open standard for connecting AI applications to outside data and services. Think of it as a plug socket: any assistant that speaks MCP can use any service that offers an MCP server, without the two sides writing custom code for each other.
There are three parts. The host is the AI app a person is using, such as ChatGPT or Claude. Inside it, a client holds the connection. The server is the service on the other end, and it offers things the assistant can use. The one that matters most for a business is tools: named actions with a description and a list of inputs, like book_meeting with a name, an email and a time. Servers can also offer resources (data to read) and prompts (ready-made instructions), but tools do the selling.
When someone asks their assistant to do something, the model reads the tool descriptions, picks one, fills in the inputs and calls it. Your server does the work and sends back a plain answer: “Booked for Tuesday at 10am” or “That slot’s gone; Wednesday at 9 is free.”
A server can run on someone’s own computer, which suits developer tools, or on the internet as a remote server, which is what a business needs. Remote servers use the Streamable HTTP transport: each message is an HTTP POST to one web address, and the reply comes back as JSON or a short stream. In practice, your MCP server is a URL, such as https://example.com/mcp, that assistants talk to.
Why a business would want one
People increasingly hand the legwork to an assistant: “find me a plumber who can come today”, “get me three quotes for a new roof”, “book a demo with a payroll tool that handles contractors”. The assistant reads a few sites, picks one, and more and more often tries to finish the job itself.
Without an MCP server, it has to work that out from your web pages. That goes wrong in familiar ways:
- Information is scattered. Opening hours in the footer, service areas on one page, prices in a PDF. The assistant may miss them or quote an old page.
- Forms are built for people. Date pickers, multi-step wizards, CAPTCHAs and “select your county” dropdowns trip agents up. Many assistants can’t submit a form at all from their servers.
- Booking widgets are worse. Embedded calendars load in frames and scripts that agents often can’t read.
- You never find out. When an assistant gives up, it recommends someone else. There’s no bounced email or abandoned-cart report.
An MCP server replaces that guesswork with a contract. The assistant knows exactly what you can do for the buyer, what details you need, and whether the request went through.
For a local service business (trades, realtors, clinics, mortgage brokers), that means the assistant can check what you offer and where, send a quote request with the job details, or book a slot on your calendar. For a B2B or SaaS company, it means accurate answers about products, plans and integrations taken from your own site, a clean demo booking with the right rep, and an inquiry that arrives with the buyer’s details filled in.
What a business MCP server offers
Most business servers need only a handful of tools. These are the common ones:
| Tool | What it does | Changes anything? |
|---|---|---|
| Business info | Who you are, what you offer, where you work, how to reach you, which forms and bookings exist | No |
| Search the website / read a page | Answers from your own pages, with links, so the assistant quotes you rather than guessing | No |
| Booking slots | Open appointment times in the buyer’s time zone | No |
| Book a meeting | Puts the buyer on your calendar; the calendar sends the confirmation | Yes |
| Send an inquiry | Delivers a message with contact details to the people who answer your contact form | Yes |
| Request a quote | Like an inquiry, with the details you need to price the job | Yes |
| Ask the business | A question in plain language, answered from your approved answers and site | No |
Each tool is described in a small piece of JSON. The description is what the model reads when deciding whether to use it, so it should say what happens and when to use it. The annotations are hints: readOnlyHint says a tool only looks things up; tools without it are treated as having real effects.
{
"name": "request_quote",
"title": "Ask for a quote",
"description": "Ask Example Plumbing for a quote. A person replies by phone or email. Only use this when the customer wants a quote, and confirm their details first.",
"inputSchema": {
"type": "object",
"properties": {
"name": { "type": "string", "description": "The customer’s full name" },
"phone": { "type": "string", "description": "Phone number, with country code outside the US" },
"service": { "type": "string", "description": "What needs doing, e.g. “replace a 50-gallon water heater”" },
"address": { "type": "string", "description": "Where the work is" }
},
"required": ["name", "phone", "service"]
},
"annotations": { "readOnlyHint": false, "destructiveHint": false, "openWorldHint": true }
}
That matters in practice. ChatGPT, for one, treats any tool without readOnlyHint as a write action and asks the user to confirm it by default. Mark lookups as read-only and they run smoothly; leave the hint off actions that reach a real person and the user gets a chance to check first.
Which assistants can connect today
As of November 2026, from each vendor’s own documentation:
| Assistant | How a remote MCP server is added |
|---|---|
| ChatGPT | On the web, in Plugins: the plus button, then “Add custom MCP server”, with the server URL and OAuth or no sign-in. Supports SSE and streaming HTTP. Workspace permissions apply; OpenAI’s page doesn’t say which plans have the option. |
| Claude | Customize, Connectors, “Add custom connector”. On Free (one custom connector), Pro, Max, Ultra, Team and Enterprise; on Team and Enterprise an owner adds it. The server must answer over HTTPS on port 443, and connections come from Anthropic’s cloud. |
| Perplexity | Plugins, “Add custom”, “Add MCP connector”, Remote. Streamable HTTP or SSE; no sign-in, API key or OAuth. In organizations, an admin has to allow members to add their own. |
| Gemini | Gemini Enterprise admins can add a custom MCP server (HTTPS, Streamable HTTP only, OAuth or no sign-in). There are reports of custom MCP apps in the consumer Gemini app on some plans, but we couldn’t confirm them in Google’s own help pages. |
| Microsoft Copilot Studio | Agent builders connect an agent to an existing MCP server and use its tools and resources. |
| Developer APIs | OpenAI’s Responses API and Anthropic’s Messages API (in beta) both accept a remote MCP server’s URL in a request, so agents that companies build on them can use your tools. |
Notice the pattern. Every one of these starts with a person, an admin or a developer typing in the server’s URL. When someone asks ChatGPT to find a roofer, it doesn’t yet look up each roofer’s MCP server by itself. That’s the honest state of things in late 2026: the plumbing for assistants to use business servers is in place across the big vendors; the plumbing for them to find one on their own is still being agreed. Our guide to how AI agents find your MCP server covers what’s settled and what you can publish now.
In the meantime, there are two other routes to the buyer. Agents that browse your site can read a <link rel="mcp"> tag or discovery files and pick the server up from there. And agents working inside a browser can use the same actions through WebMCP, the browser-side counterpart, which registers tools right on your page.
Hosted or build your own
MCP has official SDKs in several languages and packages for popular frameworks, so a basic server with one tool is an afternoon’s work for a developer. A server you’d trust with real customers is more.
| Build your own | Hosted service | |
|---|---|---|
| Hosting | A public HTTPS endpoint on port 443 that stays up, with CORS for browser clients | Run for you |
| Protocol upkeep | Track spec revisions; the July 2026 revision changed how connections work, and older clients still need supporting | The provider’s job |
| Content | Keep the server’s answers in step with your site, prices and hours | Usually synced from your site |
| Actions | Wire tools to your form handler, CRM and calendar APIs, with their own auth and quirks | Pre-built integrations |
| Abuse controls | Validation, rate limits and spam handling for every tool | Built in, to the provider’s standard |
| Discovery and analytics | Publish discovery files, log calls, attribute leads | Usually included |
| Custom logic | Anything your own code can do: live pricing, trial signup, account lookups | Limited to what the service supports |
Sign-in is the big design choice. The spec makes authorization optional. A server that only answers questions and passes on inquiries, like a contact form, can be public with no sign-in, so any assistant can use it straight away. Once a tool reads or changes a customer’s own data, such as their orders or account, you need OAuth 2.1 as the spec lays it out: a protected-resource metadata document, an authorization server, scopes, and support for the ways clients register. Each assistant then makes the user sign in to your service once. That’s worth it for a customer portal, and overkill for “send us a message”.
Building your own makes sense when your tools need your own systems, say a SaaS company that wants agents to start trials or quote live plans. For answering questions, taking inquiries and booking meetings, hosted is faster and cheaper to keep running.
Security, spam and privacy
A public MCP server is a public form with a programmable front end. The same rules apply, plus a few new ones.
- Assume anyone can call it. Not just ChatGPT on a buyer’s behalf, but scripts too. The spec says servers must validate all inputs and rate-limit tool calls. Cap submissions per caller and per day, and keep your existing spam filtering on whatever the tool feeds into.
- Make the agent confirm. The spec says hosts should keep a human in the loop and ask before sensitive operations, and the big assistants do. You can add your own guard: require a field such as “the visitor confirmed these details” on any tool that reaches a person, and say so in the description.
- Hints aren’t guarantees. Clients are told to treat annotations from servers they don’t trust as untrusted, and the same is true the other way: an agent can send you anything. Never let a tool’s description promise something your code doesn’t enforce.
- Watch what you return. Tool results go straight into the model’s context. Text you don’t control, such as reviews or user comments, can carry instructions meant for the model. Return facts, keep them short, and label third-party content.
- Expose only public facts. With no sign-in, everything a tool returns is public. Keep staff calendars, customer records and internal notes out of reach.
- Ask for little. Agents send you what buyers tell them. Ask only for what you need to reply, and tell agents to leave out personal details a search doesn’t need.
- Log every call. You’ll want to know which assistant sent a lead, spot abuse early, and show whether it’s working.
What to do now
- Pick the three to five actions that bring in money. For most businesses that’s business info, an inquiry, a quote request and booking.
- Get your facts straight. Services, areas, hours, prices or price ranges and booking rules, written down once in plain language. An MCP server is only as good as what it can tell an assistant.
- Choose hosted or built. Hosted if your tools are questions, leads and bookings; build your own if they need live data from your own systems.
- Put the same guards on it as your forms: validation, rate limits and spam filtering, with a confirmation step on anything that reaches a person.
- Connect it yourself and try it. Add the server URL as a custom connector in Claude or ChatGPT and ask for a quote, the way a customer would. You’ll see exactly what an assistant sees.
- Make it findable. Publish the discovery signals from our MCP discovery guide, add WebMCP for browser agents, and check where you stand with the free AI-readiness scan.
How Salesman Hal handles MCP
Hal gives each connected site its own hosted MCP server, at an address like https://mcp.salesmanhal.com/example.com, that any assistant can use without signing in. Its tools answer questions about the business, search and read your pages, send inquiries and quote requests, and, once you connect Calendly, Cal.com, SavvyCal, Acuity or HubSpot Meetings, show open times and book meetings. On WordPress, inquiries go through your existing Gravity Forms, Contact Form 7 or WPForms form, so your usual notifications fire. Tools that reach a person refuse to run until the agent confirms the visitor asked, and they’re rate-limited per visitor and per site. Every call is logged, so you can see which assistant sent each lead.
FAQ
Is MCP only for developers?
The protocol is technical, but the result isn’t: a web address that assistants can use to ask you questions, send you leads and book you. You don’t need to understand JSON-RPC to have one, any more than you need to understand SMTP to have an email address.
Does an MCP server help me rank in ChatGPT or Google?
Not as far as anyone has shown. No assistant documents using MCP servers as a ranking signal. Its value is that an assistant that already wants to contact or book you can actually do it. Being recommended still depends on your content, reviews, structured data and crawler access.
Will ChatGPT find my MCP server automatically?
Not today. ChatGPT, Claude, Perplexity and Gemini Enterprise all connect to a custom MCP server only after someone adds its URL. Discovery standards are settling now, and agents that browse your site can already spot a link tag or discovery file.
Do I need OAuth?
Not for public actions. The spec makes authorization optional, and a server that answers questions and passes on inquiries can work like a public contact form. Add OAuth when a tool touches a customer’s own data.
What’s the difference between MCP and WebMCP?
An MCP server runs on the internet and any assistant can call it, even one that never opens a browser. WebMCP registers tools inside your web page, for an agent working in the visitor’s browser. They complement each other.
Can spammers abuse it?
They can try, just as they do with contact forms. Rate limits per caller and per site, input validation, a confirmation step and your existing spam filters handle most of it. Log every call so you can see and block patterns.
Sources
- Specification (revision 2026-07-28), Model Context Protocol
- Transports, Model Context Protocol
- Tools, Model Context Protocol
- Authorization, Model Context Protocol
- Add custom MCP server, OpenAI
- MCP and connectors (Responses API), OpenAI
- Getting started with custom connectors using remote MCP, Claude Help Center
- MCP connector, Claude Platform docs
- Adding custom remote connectors, Perplexity Help Center
- Set up your custom MCP server data store, Google Cloud (Gemini Enterprise)
- Extend your agent with Model Context Protocol, Microsoft Learn (Copilot Studio)
Further reading
- WebMCP: a practical guide for business websites, Salesman Hal
- The MCP Registry, Model Context Protocol
Give AI assistants a way to book you.
Hal adds a hosted MCP server to your site in about ten minutes. Free during early access.
Or run the free AI-readiness scan first.