Analysis
MCP server for business: what MCP is, when it makes sense, and what WebMCP is
MCP (Model Context Protocol) is an open standard for connecting AI assistants to data and tools. A company MCP server gives assistants such as Claude, ChatGPT, Copilot, Gemini or Cursor tools that let them reach the company’s content, offer or systems directly. It does not replace SEO: the server serves people who already know the company and have added it to their assistant; it does not bring new ones. WebMCP does the same on a website — an agent in the browser gets the site’s tools instead of clicking through forms; it is still a draft, not a W3C standard.1528
AI assistants increasingly do more than answer from what they learned in training: they reach for tools — search, files, company systems. MCP has become the standard they use to do it. That raises the question of whether a business should have its own MCP server — and whether it is “the new SEO”. This article explains what MCP and WebMCP are, what an MCP server does for a business and what it does not, and how to build one securely. As an example we show the server we run on primessio.com.
What is MCP (Model Context Protocol)?
Anthropic released MCP in November 2024 as an open standard for connecting AI assistants to the systems where data lives1. The specification describes it as an open protocol for integrating LLM applications with external data sources and tools2. An MCP server can offer three kinds of things: resources (data), prompt templates and tools — functions the model can call2.
In December 2025 MCP was contributed to the Linux Foundation’s Agentic AI Foundation. The announcement says the protocol has been adopted by Claude, Cursor, Microsoft Copilot, Gemini, VS Code and ChatGPT, among others, with more than 10,000 published MCP servers5. In July 2026 the 2026-07-28 specification simplified the protocol: the session handshake is gone and every request carries its protocol version and client details3. For a business that means simpler servers that run well even on ordinary web hosting.
How widely is MCP used? There are no official statistics. A survey by the vendor Stacklok among software companies reports that nearly half use MCP in production and 19% broadly7 — a vendor study of one industry, so a signal rather than a market measure.
What an MCP server does for a business — and what it does not
An MCP server only works once a user adds it to their AI assistant: in Claude as a custom connector, including on the free plan (one there)12, in ChatGPT as a custom MCP server, subject to account and workspace policies13. From then on the assistant calls the company’s tools itself when a question needs them. That leads to the key difference from SEO:
| SEO and AEO (website content) | MCP server | WebMCP | |
|---|---|---|---|
| Who uses it | anyone searching Google or asking an AI assistant | users who added the server to their assistant | an AI agent in the browser on the company’s website |
| Does it bring new visitors | yes | no — it serves those who know the company | no — it works once the agent is already on the site |
| What the assistant gets | a page to read and cite | tools with current company data | the site’s tools instead of clicking through forms |
| Status | established channel | open standard5 | draft, Chrome trials10 |
The upshot: an MCP server does not replace work on content. It adds a channel for customers, partners and teams who work with AI assistants every day — and more accurate answers where current data matters.
When an MCP server makes sense — and when it does not
- Makes sense: when customers or partners ask about things that change — availability, order status, product documentation, terms; when your team works with an AI assistant and needs data from company systems; when you have a lot of expert content an assistant should cite with its source.
- Makes sense with limits: when tools should change something in systems — book, order, save. Then you need sign-in, narrow permissions and human confirmation before the action4.
- Does not make sense: as a replacement for SEO or the website; when content is out of date (the assistant just repeats the error faster); when nobody will measure whether the server is used.
WebMCP: website tools for agents in the browser
WebMCP is the other half of the same idea. A W3C community group draft describes an interface through which a web application provides JavaScript-based tools to AI agents8. Google presents it as a standard way of exposing structured tools so that agents can perform actions on a site with more speed, reliability and precision9. In Chrome 149 you can sign up for the origin trial10.
An important caveat: WebMCP is neither a W3C standard nor on the standards track8, and the interface name has changed between draft versions. Build it so that ordinary visitors notice nothing: the script only activates when the browser offers the interface for agents.
Security: rules before launch
The MCP specification says plainly that for trust, safety and security there should always be a human who can deny tool invocations, and that clients must treat tool descriptions as untrusted unless they come from trusted servers4. In its MCP risk list, OWASP describes tool poisoning: compromising the tools, plugins or outputs a model depends on11. For a company server that translates into a few rules:
- Read-only to start. The server returns public content and data and stores or sends nothing.
- Least privilege. Each tool does one thing; access to systems only within the task, and consequential actions only after human confirmation.
- Answers only from company data. Tools return the source text with the page URL rather than invented summaries — so the assistant can cite the source.
- Limits and logs. A rate limit per address, a call log without personal data, tests with unusual and malicious input.
- Outside content as untrusted. If the server returns external content, it keeps it separate from instructions.
Example: the Primessio MCP server
primessio.com runs a read-only MCP server at https://primessio.com/mcp. It has six tools: search across our AI analyses, articles, services and case studies; reading a full page with its sources; a list of services; the Primessio AI Implementation Framework with mandatory attribution; a qualitative check of whether a process suits a first AI implementation; and contact details. Every answer states what the site does not publish: prices, rates, delivery times and percentage promises. The same tools are available on the website through WebMCP.
The server supports both protocol generations — with the session handshake and the stateless 2026-07-28 version3 — and measurement shows which tools and questions come up and which questions the website cannot answer yet. You can add it to your AI assistant as a remote MCP server and ask, for example, how to implement AI in a business.
How to start: one data scope, read-only, measurement
- Pick a scope. Which questions from customers, partners or your team the assistant should handle, and from which data: website content, documentation, catalogue, statuses.
- Build the index from the same source as the website. Then the server and the site do not drift apart at the first content change.
- Start read-only. Tools that only return data, with the source URL in every answer.
- Add measurement. Which tools are called, what assistants ask, which questions go unanswered — the last one is a ready list of content to write.
- Publish and document. A listing in the open registry of MCP servers6, connection instructions on the website, and only after measuring, a decision on tools that change things in systems.
We design and build MCP servers and WebMCP along these rules — as part of AI implementation and automation for businesses; more on agents in our analysis of AI agents for business. The Primessio AI Implementation Framework describes the whole path, from choosing processes to continuous improvement — a method developed by Primessio in project practice.
Frequently asked questions
What is MCP?
MCP, the Model Context Protocol, is an open protocol that enables seamless integration between LLM applications and external data sources and tools. Anthropic released it in November 2024 as a standard for connecting AI assistants to the systems where data lives and contributed it to the Linux Foundation’s Agentic AI Foundation in December 2025.215
What is an MCP server?
A service that offers an AI assistant resources, prompt templates and tools — functions the model can call. A company MCP server can, for example, search its content, describe its services or return an order status. The assistant calls these tools itself when a question needs them.2
Will an MCP server replace SEO?
No. An assistant only uses an MCP server once the user adds it — in Claude as a custom connector, in ChatGPT as a custom MCP server. New people still find the company through search and AI answers based on its website content. An MCP server is an extra channel for those who already know the company.1213
What is WebMCP?
WebMCP is an interface through which a website provides JavaScript-based tools to AI agents in the browser. Instead of clicking through forms, the agent calls a ready-made function of the site. It is a W3C community group draft, not a W3C standard; Chrome offers it in an origin trial.8810
Is an MCP server secure?
As secure as its permissions. The specification says there should always be a human who can deny tool invocations, and OWASP describes risks such as tool poisoning — compromising the tools or outputs a model depends on. That is why we start with a read-only server, public data, a rate limit and a call log.411
Sources
- Anthropic: Introducing the Model Context Protocol
- Model Context Protocol (Linux Foundation / Agentic AI Foundation): Model Context Protocol Specification, version 2026-07-28
- Model Context Protocol (Linux Foundation / Agentic AI Foundation): Key Changes (2026-07-28)
- Model Context Protocol (Linux Foundation / Agentic AI Foundation): Tools (Model Context Protocol Specification)
- The Linux Foundation: Linux Foundation Announces the Formation of the Agentic AI Foundation
- Model Context Protocol Blog: Introducing the MCP Registry
- Stacklok: State of Model Context Protocol in Software 2026
- W3C Web Machine Learning Community Group: WebMCP (Draft Community Group Report, 20 October 2026)
- Google Chrome for Developers: WebMCP is available for early preview
- Google Chrome for Developers: Join the WebMCP origin trial
- OWASP Foundation: OWASP MCP Top 10 (beta)
- Anthropic Help Center: Get started with custom connectors using remote MCP
- OpenAI: Connect and test your plugin