On this page
Every few years a connector wins and the drawer of adapters goes quiet. MCP is that moment for AI tools — and unusually for a technical standard, the analogy in its own documentation is the whole explanation.
What MCP is, in one analogy
"Think of MCP like a USB-C port for AI applications. Just as USB-C provides a standardized way to connect electronic devices, MCP provides a standardized way to connect AI applications to external systems" (Model Context Protocol docs). It is open, and it is supported across the assistants and editors your customers and staff already use — Claude, ChatGPT, VS Code, Cursor among them.
Before MCP, every pairing of an assistant and a business system was a bespoke integration. After it, a system that speaks MCP once can be used by any assistant that speaks MCP — including the one on your website.
What it changes for a customer
The difference is between an assistant that describes and one that does.
| Without a connection | With your systems connected |
|---|---|
| "You can check your order on the orders page." | "Order 4182 shipped Tuesday, arriving Friday." |
| "Our opening hours are on the contact page." | "The 3pm Thursday slot is free — shall I book it?" |
| "Please email us to change your address." | "I've updated it — here is what I changed." |
Nothing in that right-hand column is model cleverness. It is your booking system, your order table, your CRM, reachable through one standard shape.
Three access levels and a confirmation card
Connecting systems to an assistant is where owners are right to be careful, so the controls are deliberately boring:
- Open to anyone — safe, non-personal look-ups, for visitors who have not signed in.
- Signed-in customers — look-ups and actions on their own data only.
- On the customer's behalf — the same, for actions you want the customer to authorise explicitly.
On top of those, any change can be marked as needing a confirmation card: Your mate shows the exact action and does nothing until the customer says yes. Your server works out who the customer is only from a verified signed token, and returns only that person's data — never from an account id in a tool argument. The levels are laid out in connecting your systems, and the reasoning behind the confirmation step is on our security page.
Adding one, without writing the integration
If your systems already speak MCP, there is no platform-specific glue to write. The path is four steps, and the connect an MCP server guide has the detail:
1. Serve tools/list over HTTPS at a stable URL. https://[your-site]/mcp
2. Add the connection and run the probe — it lists
every tool it finds.
3. Set each tool's access level, and tick Confirm on
every change. [CHECK] mark deliberately
4. Run the checks and publish. A failed probe blocks
the release instead of shipping half a connection.A tool you forget to mark as a change will run without a card, so step 3 is the one to do slowly. You can test a server before attaching it with MCP check.
Ask your AI assistant to do it
The management surface is itself an MCP server, which means the setup can be done conversationally:
Probe my MCP server at https://[your-site]/mcp and list every tool it exposes.
For each tool, propose an access level — open to anyone, signed-in customers, or
on the customer's behalf — and say whether it changes data. Mark every change as
needing confirmation. Show me the full table for approval before saving anything,
and flag anything you cannot classify with [CHECK].Look-ups run right away; the write shows its payload and waits for your yes. More of that pattern is in managing from chat.
MCP and WebMCP are not the same thing
Worth keeping straight, because the names invite confusion. MCP connects an assistant to your systems, behind the scenes. WebMCP lets your web page publish its own actions to whatever agent is browsing it — ChatGPT's desktop browser already calls tools registered this way (ChatGPT docs). One serves the assistant you run; the other serves the agents that visit. The page tools guide covers the second, and machine-readable context such as our own /llms.txt helps both.
Most owners should start smaller than any of this: build an assistant from your URL, watch what customers actually ask it, and connect a system when a question arrives that only your systems can answer. Pricing starts at that scale, not at the enterprise one.
Questions
What does MCP stand for?
Model Context Protocol — an open standard for connecting AI applications to external systems. Its own documentation describes it as a USB-C port for AI applications, a single shape that fits many tools.
Do I need MCP if I only want an assistant that answers questions?
No. Answering from your published pages needs no connection at all. MCP is what you add when you want the assistant to look something up in your systems or do something on a customer's behalf.
Is MCP the same thing as WebMCP?
No. MCP connects an assistant to a server behind the scenes. WebMCP lets a web page publish its own actions to whatever agent is browsing it. They solve different halves of the same problem.
How does my server know which customer is asking?
From a verified bearer token it receives with the call — either a short-lived signed token or a per-customer OAuth token. Never from an account id passed as a tool argument, which anyone could type.
What if I do not have an MCP server?
Then you are in the majority, and you may not need one yet. Start with an assistant that answers from your site; add a connection when a real question keeps arriving that only your systems can answer.